Summary: Previously if we had a bitcast vector output type that needs promotion and a vector input type that needs widening we would just do a stack store and load to handle the conversion. We can do a little better if we can widen the bitcast to a legal vector type the same size as the widened input type. Then we can do the bitcast between this widened type and the widened input type. Afterwards we can extract_subvector back to the original output and any_extend that. Type legalization will then circle back and handle promotion of the extract_subvector and the any_extend will just be removed. This will avoid going through the stack and allows us to remove a custom version of this legalization from X86.
Reviewers: efriedma, RKSimon
Reviewed By: efriedma
Subscribers: javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D53229
llvm-svn: 345567
Use SelectionDAG::EVTToAPFloatSemantics. Make the LogicVT calculation in LowerFABSorFNEG similar to LowerFCOPYSIGN. Use APInt::getSignedMaxValue instead of ~APInt::getSignMask.
llvm-svn: 345565
The machine verifier was disabled for x86 by default. There are now only
9 tests failing, compared to what previously was between 20 and 30.
This is a good opportunity to file bugs for all the remaining issues,
then explicitly disable the failing tests and enabling the machine
verifier by default.
This allows us to avoid adding new tests that break the verifier.
PR27481
llvm-svn: 345513
When the floating point constants are whole numbers they have no decimal point so look like integers, but mean something very different in something like an 'and' instruction.
Ideally we would just print a decimal point and a 0, but I couldn't see how to make APFloat::toString do that.
llvm-svn: 345488
Add vector support to TargetLowering::expandFP_TO_UINT.
This exposes an issue in X86TargetLowering::LowerVSELECT which was assuming that the select mask was the same width as the LHS/RHS ops - as long as the result is a sign splat we can easily sext/trunk this.
llvm-svn: 345473
Makes no difference to actual shuffle decoding yet, but merges all the existing limits in one place for when proper support is fixed.
........
Its been reported that this is causing out of trunk failures.
llvm-svn: 345451
Summary:
The main challenge here is that X86InstrInfo::AnalyzeBranch doesn't
understand the way we're using a CALL instruction as a branch, so we
can't list the CallTarget MBB as a successor of the entry block. If we
don't list it as a successor, then the AsmPrinter doesn't print a label
for the MBB.
Fix the issue by inserting our own label at the beginning of the call
target block. We can rely on the AsmPrinter to always emit it, even
though the block appears to be unreachable, but address-taken.
Fixes PR38391.
Reviewers: thegameg, chandlerc, echristo
Subscribers: hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D53653
llvm-svn: 345426
These promotions add additional bitcasts to the SelectionDAG that can pessimize computeKnownBits/computeNumSignBits. It also seems to interfere with broadcast formation.
This patch removes the promotion and adds isel patterns instead.
The increased table size is more than I would like, but hopefully we can find some canonicalizations or other tricks to start pruning out patterns going forward.
Differential Revision: https://reviews.llvm.org/D53268
llvm-svn: 345408
This is a narrow fix for 1 of the problems mentioned in PR27780:
https://bugs.llvm.org/show_bug.cgi?id=27780
I looked at more general solutions, but it's a mess. We canonicalize shuffle masks
based on the number of elements accessed from each operand, and that's not optional.
If you remove that, we'll crash because we fail to match isel patterns. So I'm
waiting until we're sure that we have blendvb with constant condition and then
commuting based on the load potential. Other cases like blend-with-immediate are
already handled elsewhere, so this is probably not a common problem anyway.
I didn't use "MayFoldLoad" because that checks for one-use and in these cases, we've
screwed that up by creating a temporary PSHUFB using these operands that we're counting
on to be killed later. Undoing that didn't look like a simple task because it's
intertwined with determining if we actually use both operands of the shuffle or not.a
Differential Revision: https://reviews.llvm.org/D53737
llvm-svn: 345390
The required-vector-width attribute was only used for backend testing and has never been generated by clang.
I believe clang is now generating min-legal-vector-width for vector uses in user code.
With this I believe passing -mprefer-vector-width=256 to clang should prevent use of zmm registers in the generated assembly unless the user used a 512-bit intrinsic in their source code.
llvm-svn: 345317
KNL is based on a modified Silvermont core so I don't think these features apply. I think the LEA flag is probably also wrong, but I'm less sure as I barely understand the 3 LEA flags we have currently.
Differential Revision: https://reviews.llvm.org/D53671
llvm-svn: 345285
Summary:
The pfm counters are now in the ExegesisTarget rather than the
MCSchedModel (PR39165).
This also compresses the pfm counter tables (PR37068).
Reviewers: RKSimon, gchatelet
Subscribers: mgrang, llvm-commits
Differential Revision: https://reviews.llvm.org/D52932
llvm-svn: 345243
Multiply a is complex operation so just because some bit of the output isn't used doesn't mean that bit of the input isn't used.
We might able to bound it, but it will require some more thought.
llvm-svn: 345241
Instead of using the MOVGOT64r pseudo, use the existing
MO_PIC_BASE_OFFSET support on symbol operands. Now I don't have to
create a "scratch register operand" for the pseudo to use, and the
register allocator can make better decisions.
Fixes some X86 verifier errors tracked in PR27481.
llvm-svn: 345219
It's possible to do a tail call to a stack argument. LLVM already
calculates the right stack offset to call through.
Fixes the sibcall* and musttail* verifier failures tracked at PR27481.
llvm-svn: 345197
Summary:
This renames the IsParsingMSInlineAsm member variable of AsmLexer to
LexMasmIntegers and moves it up to MCAsmLexer. This is the only behavior
controlled by that variable. I added a public setter, so that it can be
set from outside or from the llvm-mc command line. We may need to
arrange things so that users can get this behavior from clang, but
that's future work.
I also put additional hex literal lexing functionality under this flag
to fix PR32973. It appears that this hex literal parsing wasn't intended
to be enabled in non-masm-style blocks.
Now, masm integers (0b1101 and 0ABCh) work in __asm blocks from clang,
but 0b label references work when using .intel_syntax in standalone .s
files.
However, 0b label references will *not* work from __asm blocks in clang.
They will work from GCC inline asm blocks, which it sounds like is
important for Crypto++ as mentioned in PR36144.
Essentially, we only lex masm literals for inline asm blobs that use
intel syntax. If the .intel_syntax directive is used inside a gnu-style
inline asm statement, masm literals will not be lexed, which is
compatible with gas and llvm-mc standalone .s assembly.
This fixes PR36144 and PR32973.
Reviewers: Gerolf, avt77
Subscribers: eraman, hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D53535
llvm-svn: 345189
I'm not sure all the microarchitectural tuning flags that have been added to IVBFeatures are relevant for KNL. Separating will allow us to see and audit them. There might even be some simplification opportunities in the Sandy Bridge through Icelake inheritance line without KNL using the same chain.
llvm-svn: 345183
Add X86 SimplifyDemandedBitsForTargetNode and use it to simplify PMULDQ/PMULUDQ target nodes.
This enables us to repeatedly simplify the node's arguments after the previous approach had to be reverted due to PR39398.
Differential Revision: https://reviews.llvm.org/D53643
llvm-svn: 345182
This patch brings back the MOV64r0 pseudo instruction for zeroing a 64-bit register. This replaces the SUBREG_TO_REG MOV32r0 sequence we use today. Post register allocation we will rewrite the MOV64r0 to a 32-bit xor with an implicit def of the 64-bit register similar to what we do for the various XMM/YMM/ZMM zeroing pseudos.
My main motivation is to enable the spill optimization in foldMemoryOperandImpl. As we were seeing some code that repeatedly did "xor eax, eax; store eax;" to spill several registers with a new xor for each store. With this optimization enabled we get a store of a 0 immediate instead of an xor. Though I admit the ideal solution would be one xor where there are multiple spills. I don't believe we have a test case that shows this optimization in here. I'll see if I can try to reduce one from the code were looking at.
There's definitely some other machine CSE(and maybe other passes) behavior changes exposed by this patch. So it seems like there might be some other deficiencies in SUBREG_TO_REG handling.
Differential Revision: https://reviews.llvm.org/D52757
llvm-svn: 345165
Non-uniform division/remainder handling was added back at D49248/D50765 - so share the 'mul+sub' costs that already exist for uniform cases.
llvm-svn: 345164
This B/W VPTEST instructions are only available with AVX512BW. But lowering should prevent any byte or word elements from getting to isel so this can't be exposed.
llvm-svn: 345112
When implementing memset's today we often see this pattern:
$x0 = MOV 0xXYXYXYXYXYXYXYXY
store $x0, ...
$w1 = MOV 0xXYXYXYXY
store $w1, ...
We first create a 64bit constant in a 64bit register with all bytes the
same and then create a 32bit constant with all bytes the same in a 32bit
register. In many targets we could just access the lower byte of the
64bit register instead.
- Ideally this would be handled by the ConstantHoist pass but it runs
too early when memset isn't expanded yet.
- The memset expansion code already had this optimization implemented,
however SelectionDAG constantfolding would constantfold the
"trunc(bigconstnat)" pattern to "smallconstant".
- This patch makes the memset expansion mark the constant as Opaque and
stop DAGCombiner from constant folding in this situation. (Similar to
how ConstantHoisting marks things as Opaque to avoid folding
ADD/SUB/etc.)
Differential Revision: https://reviews.llvm.org/D53181
llvm-svn: 345102
We can't add the MULDQ node back to the worklist after the demanded bits change has been committed in case the node has been removed entirely. This will have to wait until we have SimplifyDemandedBitsForTargetNode.
llvm-svn: 345070
This initially landed in rL345014, but was reverted in rL345017
due to sanitizer-x86_64-linux-fast buildbot failure in
check-lld (ELF/relocatable-versioned.s) test.
While i'm not yet quite sure what is the problem, one obvious
thing here is that extra truncation roundtrip.
Maybe that's it? If not, will re-revert.
Differential Revision: https://reviews.llvm.org/D53521
llvm-svn: 345027
Matches the approach taken in the constant pool shuffle decoders, and uses an UndefElts mask instead of uint64_t(-1) raw mask values, which doesn't work safely for i32/i64 shuffle mask sizes (as the -1 value is legal).
This allows us to remove the constant pool shuffle decoders from most of the getTargetShuffleMask variable shuffle cases (X86ISD::VPERMV3 will be handled in a future commit).
llvm-svn: 345018
Summary:
Continuation of D52348.
We also get the `c) x & (-1 >> (32 - y))` pattern here, because of the D48768.
I will add extra-uses into those tests and follow-up with a patch to handle those patterns too.
Reviewers: RKSimon, craig.topper
Reviewed By: craig.topper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D53521
llvm-svn: 345014
analyzeBranch()/insertBranch() etc. do not properly deal with an undef
flag on the eflags input and used to produce invalid MIR. I don't see
this ever affecting real world inputs (I don't think it is possible to
produce undef flags with llvm IR), so I simply changed the code to bail
out in this case.
rdar://42122367
llvm-svn: 344970
I've included a fix to DAGCombiner::ForwardStoreValueToDirectLoad that I believe will prevent the previous miscompile.
Original commit message:
Theoretically this was done to simplify the amount of isel patterns that were needed. But it also meant a substantial number of our isel patterns have to match an explicit bitcast. By making the vXi32/vXi16/vXi8 types legal for loads, DAG combiner should be able to change the load type to rem
I had to add some additional plain load instruction patterns and a few other special cases, but overall the isel table has reduced in size by ~12000 bytes. So it looks like this promotion was hurting us more than helping.
I still have one crash in vector-trunc.ll that I'm hoping @RKSimon can help with. It seems to relate to using getTargetConstantFromNode on a load that was shrunk due to an extract_subvector combine after the constant pool entry was created. So we end up decoding more mask elements than the lo
I'm hoping this patch will simplify the number of patterns needed to remove the and/or/xor promotion.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits, RKSimon
Differential Revision: https://reviews.llvm.org/D53306
llvm-svn: 344965
A while ago we changed pushf and popf in Intel mode to generate pushfq
and popfq. Unfortunately that left us with no way to get the 16-bit
encoding in Intel mode so this patch adds pushfw and popfw as aliases
there.
llvm-svn: 344949
We can't safely assume that certain RawMask entries are UNDEF as most variable shuffles ignore non-index bits - PSHUFB only works on i8 elts so it'd be safe to use but I'm intending to come up with an alternative approach that works for all.
........
Enable this for PSHUFB constant mask decoding and remove the ConstantPool DecodePSHUFBMask
llvm-svn: 344937
We can't safely assume that certain RawMask entries are UNDEF as most variable shuffles ignore non-index bits.
........
Add support for UNDEF raw mask elements and remove the ConstantPool DecodeVPERMILPMask usage in X86ISelLowering.cpp
llvm-svn: 344936
Summary:
As discussed in D52304 / IRC, we now have pattern matching for
'bit extract' in two places - tablegen and `X86DAGToDAGISel`.
There are 4 patterns.
And we will have a problem with `x & (-1 >> (32 - y))` pattern.
* If the mask is one-use, then it is always unfolded into `x << (32 - y) >> (32 - y)` first.
Thus, the existing test coverage is already broken.
* If it is not one-use, then it is not unfolded, and is matched as BZHI.
* If it is not one-use, we will not match it as BEXTR. And if it is one-use, it will have been unfolded already.
So we will either not handle that pattern for BEXTR, or not have test coverage for it.
This is bad.
As discussed with @craig.topper, let's unify this matching, and do everything in `X86DAGToDAGISel`.
Then we will not have code duplication, and will have proper test coverage.
This indeed does not affect any tests, and this is great.
It means that for these two patterns, the `X86DAGToDAGISel` is identical to the tablegen version.
Please review carefully, i'm not fully sure about that intrinsic change, and introduction of the new `X86ISD` opcode.
Reviewers: craig.topper, RKSimon, spatel
Reviewed By: craig.topper
Subscribers: llvm-commits, craig.topper
Differential Revision: https://reviews.llvm.org/D53164
llvm-svn: 344904
Summary:
Trivial continuation of D52304.
While this pattern is not canonical, we do select it in the BZHI case,
so this should not be any different.
Reviewers: RKSimon, craig.topper, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D52348
llvm-svn: 344902
This makes fast isel treat all legal vector types the same way. Previously only vXi64 was in the fast-isel tables.
This unfortunately prevents matching of andn by fast-isel for these types since the requires SelectionDAG. But we already had this issue for vXi64. So at least we're consistent now.
Interestinly it looks like fast-isel can't handle instructions with constant vector arguments so the the not part of the andn patterns is selected with SelectionDAG. This explains why VPTERNLOG shows up in some of the tests.
This is a subset of D53268. As I make progress on that, I will try to reduce the number of lines in the tablegen files.
llvm-svn: 344884
Summary:
Theoretically this was done to simplify the amount of isel patterns that were needed. But it also meant a substantial number of our isel patterns have to match an explicit bitcast. By making the vXi32/vXi16/vXi8 types legal for loads, DAG combiner should be able to change the load type to remove the bitcast.
I had to add some additional plain load instruction patterns and a few other special cases, but overall the isel table has reduced in size by ~12000 bytes. So it looks like this promotion was hurting us more than helping.
I still have one crash in vector-trunc.ll that I'm hoping @RKSimon can help with. It seems to relate to using getTargetConstantFromNode on a load that was shrunk due to an extract_subvector combine after the constant pool entry was created. So we end up decoding more mask elements than the load size.
I'm hoping this patch will simplify the number of patterns needed to remove the and/or/xor promotion.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits, RKSimon
Differential Revision: https://reviews.llvm.org/D53306
llvm-svn: 344877
Summary:
These nodes exist to overcome an isel problem where we can generate a zero extend of an AH register followed by an extract subreg, and another zero extend. The first zero extend exists to avoid a partial register update copying the AH register into the low 8-bits. The second zero extend exists if the user wanted the remainder zero extended.
To make this work we had a DAG combine to morph the DIVREM opcode to a special opcode that included the extend. But then we had to add the new node to computeKnownBits and computeNumSignBits to process the extension portion.
This patch instead removes all of that and adds a late peephole to detect the two extends.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D53449
llvm-svn: 344874
D53306 exposes an issue where we sometimes use constant pool data from bigger vectors than the target shuffle mask. This should be safe to do, but we have to be certain that we're using the bottom most part of the vector as the shuffle mask decoders have no way to peek into subvectors with non-zero offsets.
llvm-svn: 344867
There is no guarantee the root is at the end if isel created any nodes without morphing them. This includes the nodes created by manual isel from C++ code in X86ISelDAGToDAG.
This is similar to r333415 from PowerPC which is where I originally stole the peephole loop from.
I don't have a test case, but without this a future patch doesn't work which is how I found it.
llvm-svn: 344808
Allows to disable direct TLS segment access (%fs or %gs). GCC supports
a similar flag, it can be useful in some circumstances, e.g. when a thread
context block needs to be updated directly from user space. More info
and specific use cases: https://bugs.llvm.org/show_bug.cgi?id=16145
There is another revision for clang as well.
Related: D53102
All X86 CodeGen tests appear to pass:
```
[46/47] Running lit suite /SourceCache/llvm-trunk-8.0/test/CodeGen
Testing Time: 23.17s
Expected Passes : 3801
Expected Failures : 15
Unsupported Tests : 8021
```
Reviewed by: Craig Topper.
Patch by nruslan (Ruslan Nikolaev).
Differential Revision: https://reviews.llvm.org/D53103
llvm-svn: 344723
Without this we match the CMP+AND to a TEST and then match the SHR separately. I'm trusting analyzeCompare to remove the TEST during the peephole pass. Otherwise we need to check the flag users to see if they only use the Z flag.
This recovers a case lost by r344270.
Differential Revision: https://reviews.llvm.org/D53310
llvm-svn: 344649
These included a bitcast of a load from v4f32 to v2f64, but DAG combine should have already changed the type of the load to remove the cast.
llvm-svn: 344573
by `getTerminator()` calls instead be declared as `Instruction`.
This is the biggest remaining chunk of the usage of `getTerminator()`
that insists on the narrow type and so is an easy batch of updates.
Several files saw more extensive updates where this would cascade to
requiring API updates within the file to use `Instruction` instead of
`TerminatorInst`. All of these were trivial in nature (pervasively using
`Instruction` instead just worked).
llvm-svn: 344502
Summary:
I've noticed that the bitcasts we introduce for these make computeKnownBits and computeNumSignBits not work well in LegalizeVectorOps. LegalizeVectorOps legalizes bottom up while LegalizeDAG legalizes top down. The bottom up strategy for LegalizeVectorOps means operands are legalized before their uses. So we promote and/or/xor before we legalize the operands that use them making computeKnownBits/computeNumSignBits in places like LowerTruncate suboptimal. I looked at changing LegalizeVectorOps to be top down as well, but that was more disruptive and caused some regressions. I also looked at just moving promotion of binops to LegalizeDAG, but that had a few issues one around matching AND,ANDN,OR into VSELECT because I had to create ANDN as vXi64, but the other nodes hadn't legalized yet, I didn't look too hard at fixing that.
This patch seems to produce better results overall than my other attempts. We now form broadcasts of constants better in some cases. For at least some of them the AND was being introduced in LegalizeDAG, promoted to vXi64, and the BUILD_VECTOR was also legalized there. I think we got bad ordering of that. Now the promotion is out of the legalizer so we handle this better.
In the longer term I think we really should evaluate whether we should be doing this promotion at all. It's really there to reduce isel pattern count, but I'm wondering if we'd be better served just eating the pattern cost or doing C++ based isel for vector and/or/xor in X86ISelDAGToDAG. The masked and/or/xor will definitely be difficult in patterns if a bitcast gets between the vselect and the and/or/xor node. That becomes a lot of permutations to cover.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D53107
llvm-svn: 344487
interleave-group
The vectorizer currently does not attempt to create interleave-groups that
contain predicated loads/stores; predicated strided accesses can currently be
vectorized only using masked gather/scatter or scalarization. This patch makes
predicated loads/stores candidates for forming interleave-groups during the
Loop-Vectorizer's analysis, and adds the proper support for masked-interleave-
groups to the Loop-Vectorizer's planning and transformation stages. The patch
also extends the TTI API to allow querying the cost of masked interleave groups
(which each target can control); Targets that support masked vector loads/
stores may choose to enable this feature and allow vectorizing predicated
strided loads/stores using masked wide loads/stores and shuffles.
Reviewers: Ayal, hsaito, dcaballe, fhahn, javed.absar
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D53011
llvm-svn: 344472
Summary: This is similar to what D52528 did for loads. It should match what generic type legalization does in 64-bit mode where it uses a v2i64 cast and an i64 store.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D53173
llvm-svn: 344470
There is one remnant - AVX1 custom splitting of 256-bit vectors - which is due to a regression where the X86ISD::ANDNP is still performed as a YMM.
I've also tightened the CTLZ or CTPOP lowering in SelectionDAGLegalize::ExpandBitCount to require a legal CTLZ - it doesn't affect existing users and fixes an issue with AVX512 codegen.
llvm-svn: 344457
Use isConstantSplat instead of ISD::isConstantSplatVector to let us us peek through to illegal types (in this case for i686 targets to recognise i64 constants)
llvm-svn: 344452
If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen.
Similar to rL344447, this is also closer to LegalizeDAG's approach
llvm-svn: 344448
This patch changes the vector CTTZ lowering from:
cttz(x) = ctpop((x & -x) - 1)
to:
cttz(x) = ctpop(~x & (x - 1))
Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first).
Differential Revision: https://reviews.llvm.org/D53214
llvm-svn: 344447
Add shuffle lowering for the case where we can shuffle the lanes into place followed by an in-lane permute.
This is mainly for cases where we can have non-repeating permutes in each lane, but for now I've just enabled it for v4f64 unary shuffles to fix PR39161 - there is no test coverage for other shuffles that might benefit yet.
We now have several cross-lane shuffle lowering methods that all do something similar - I've looked at merging some of these (notably by making the repeated mask mechanism in lowerVectorShuffleByMerging128BitLanes optional), but there is a lot of assertions/assumptions in the way that makes this tricky - I ended up going for adding yet another relatively simple method instead.
Differential Revision: https://reviews.llvm.org/D53148
llvm-svn: 344446
Generic legalization should be able to finish legalizing the EXTRACT_SUBVECTOR probably by turning it into a BUILD_VECTOR. But we should emit the simplest sequence.
llvm-svn: 344424
The algorithm we would do previously was identical to generic legalization. If we ever switch to legalizing integer vectors via widening we'll be able to kill off the code since it now only runs for promotion.
llvm-svn: 344423
This is the planned follow-up to D52997. Here we are reducing horizontal vector math codegen
by default. AMD Jaguar (btver2) should have no difference with this patch because it has
fast-hops. (If we want to set that bit for other CPUs, let me know.)
The code changes are small, but there are many test diffs. For files that are specifically
testing for hops, I added RUNs to distinguish fast/slow, so we can see the consequences
side-by-side. For files that are primarily concerned with codegen other than hops, I just
updated the CHECK lines to reflect the new default codegen.
To recap the recent horizontal op story:
1. Before rL343727, we were producing hops for all subtargets for a variety of patterns.
Hops were likely not optimal for all targets though.
2. The IR improvement in r343727 exposed a hole in the backend hop pattern matching, so
we reduced hop codegen for all subtargets. That was bad for Jaguar (PR39195).
3. We restored the hop codegen for all targets with rL344141. Good for Jaguar, but
probably bad for other CPUs.
4. This patch allows us to distinguish when we want to produce hops, so everyone can be
happy. I'm not sure if we have the best predicate here, but the intent is to undo the
extra hop-iness that was enabled by r344141.
Differential Revision: https://reviews.llvm.org/D53095
llvm-svn: 344361
Pull out repeated byte sum stage for popcount of vector elements > 8bits.
This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw.
llvm-svn: 344348
Fixes PR32160 by reducing the size of PSHUFB if we only use one of the lanes.
This approach can probably be generalized to handle any target shuffle (and any subvector index) but we have no test coverage at the moment.
llvm-svn: 344336
This patch adds the ability to identify instructions that are "move elimination
candidates". It also allows scheduling models to describe processor register
files that allow move elimination.
A move elimination candidate is an instruction that can be eliminated at
register renaming stage.
Each subtarget can specify which instructions are move elimination candidates
with the help of tablegen class "IsOptimizableRegisterMove" (see
llvm/Target/TargetInstrPredicate.td).
For example, on X86, BtVer2 allows both GPR and MMX/SSE moves to be eliminated.
The definition of 'IsOptimizableRegisterMove' for BtVer2 looks like this:
```
def : IsOptimizableRegisterMove<[
InstructionEquivalenceClass<[
// GPR variants.
MOV32rr, MOV64rr,
// MMX variants.
MMX_MOVQ64rr,
// SSE variants.
MOVAPSrr, MOVUPSrr,
MOVAPDrr, MOVUPDrr,
MOVDQArr, MOVDQUrr,
// AVX variants.
VMOVAPSrr, VMOVUPSrr,
VMOVAPDrr, VMOVUPDrr,
VMOVDQArr, VMOVDQUrr
], CheckNot<CheckSameRegOperand<0, 1>> >
]>;
```
Definitions of IsOptimizableRegisterMove from processor models of a same
Target are processed by the SubtargetEmitter to auto-generate a target-specific
override for each of the following predicate methods:
```
bool TargetSubtargetInfo::isOptimizableRegisterMove(const MachineInstr *MI)
const;
bool MCInstrAnalysis::isOptimizableRegisterMove(const MCInst &MI, unsigned
CPUID) const;
```
By default, those methods return false (i.e. conservatively assume that there
are no move elimination candidates).
Tablegen class RegisterFile has been extended with the following information:
- The set of register classes that allow move elimination.
- Maxium number of moves that can be eliminated every cycle.
- Whether move elimination is restricted to moves from registers that are
known to be zero.
This patch is structured in three part:
A first part (which is mostly boilerplate) adds the new
'isOptimizableRegisterMove' target hooks, and extends existing register file
descriptors in MC by introducing new fields to describe properties related to
move elimination.
A second part, uses the new tablegen constructs to describe move elimination in
the BtVer2 scheduling model.
A third part, teaches llm-mca how to query the new 'isOptimizableRegisterMove'
hook to mark instructions that are candidates for move elimination. It also
teaches class RegisterFile how to describe constraints on move elimination at
PRF granularity.
llvm-mca tests for btver2 show differences before/after this patch.
Differential Revision: https://reviews.llvm.org/D53134
llvm-svn: 344334
DIV/REM by constants should always be expanded into mul/shift/etc.
patterns. Unfortunately the ConstantHoisting pass runs too early at a
point where the pattern isn't expanded yet. However after
ConstantHoisting hoisted some immediate the result may not expand
anymore. Also the hoisting typically doesn't make sense because it
operates on immediates that will change completely during the expansion.
Report DIV/REM as TCC_Free so ConstantHoisting will not touch them.
Differential Revision: https://reviews.llvm.org/D53174
llvm-svn: 344315
On 64-bit targets the generic legalize will use an i64 load and a scalar_to_vector for us. But on 32-bit targets i64 isn't legal and the generic legalizer will end up emitting two 32-bit loads. We have DAG combines that try to put those two loads back together with pretty good success.
This patch instead uses f64 to avoid the splitting entirely. I've made it do the same for 64-bit mode for consistency and to keep the load in the fp domain.
There are a few things in here that look like regressions in 32-bit mode, but I believe they bring us closer to the 64-bit mode codegen. And that the 64-bit mode code could be better. I think those issues should be looked at separately.
Differential Revision: https://reviews.llvm.org/D52528
llvm-svn: 344291
This is an alternative to D53080 since I think using a BEXTR for a shifted mask is definitely an improvement when the shl can be absorbed into addressing mode. The other cases I'm less sure about.
We already have several tricks for handling an and of a shift in address matching. This adds a new case for BEXTR.
I've moved the BEXTR matching code back to X86ISelDAGToDAG to allow it to match. I suppose alternatively we could directly emit a X86ISD::BEXTR node that isel could pattern match. But I'm trying to view BEXTR matching as an isel concern so DAG combine can see 'and' and 'shift' operations that are well understood. We did lose a couple cases from tbm_patterns.ll, but I think there are ways to recover that.
I've also put back the manual load folding code in matchBEXTRFromAnd that I removed a few months ago in r324939. This gives us some more freedom to make decisions based on the ability to fold a load. I haven't done anything with that yet.
Differential Revision: https://reviews.llvm.org/D53126
llvm-svn: 344270
Summary:
As discussed in D48491, we can't really do this in the TableGen,
since we need to produce *two* instructions. This only implements
one single pattern. The other 3 patterns will be in follow-ups.
I'm not sure yet if we want to also fuse shift into here
(i.e `(x >> start) & ...`)
Reviewers: RKSimon, craig.topper, spatel
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D52304
llvm-svn: 344224
Remove tryFoldVecLoad since tryFoldLoad would call IsProfitableToFold and pick up the new check.
This saves about 5K out of ~600K on the generated isel table.
llvm-svn: 344189
Summary:
As discussed in [[ https://bugs.llvm.org/show_bug.cgi?id=38938 | PR38938 ]],
we fail to emit `BEXTR` if the mask is shifted.
We can't deal with that in `X86DAGToDAGISel` `before the address mode for the inc is selected`,
and we can't really do it in the normal DAGCombine, because we don't have generic `ISD::BitFieldExtract` node,
and if we simply turn the shifted mask into a normal mask + shift-left, it will be folded back.
So it would seem X86ISelLowering is the place to handle this.
This patch only moves the matchBEXTRFromAnd()
from X86DAGToDAGISel to X86ISelLowering.
It does not add support for the 'shifted mask' pattern.
Reviewers: RKSimon, craig.topper, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D52426
llvm-svn: 344179
This is intended to restore horizontal codegen to what it looked like before IR demanded elements improved in:
rL343727
As noted in PR39195:
https://bugs.llvm.org/show_bug.cgi?id=39195
...horizontal ops can be worse for performance than a shuffle+regular binop, so I've added a TODO. Ideally, we'd
solve that in a machine instruction pass, but a quicker solution will be adding a 'HasFastHorizontalOp' feature
bit to deal with it here in the DAG.
Differential Revision: https://reviews.llvm.org/D52997
llvm-svn: 344141
Similar to what already happens in the DAGCombiner wrappers, this patch adds the root nodes back onto the worklist if the DCI wrappers' SimplifyDemandedBits/SimplifyDemandedVectorElts were successful.
Differential Revision: https://reviews.llvm.org/D53026
llvm-svn: 344132
This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step.
Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast.
There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it.
This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues.
llvm-svn: 344071
As noted in D52747, if we prefer IR to use trunc for bool vectors rather
than and+icmp, we can expose codegen shortcomings as seen here with masked store.
Replace a hard-coded PCMPGT simplification with the more general demanded bits call
to improve things.
Differential Revision: https://reviews.llvm.org/D52964
llvm-svn: 344048
As discussed on D52964, this adds 256-bit *_EXTEND_VECTOR_INREG lowering support for AVX1 targets to help improve SimplifyDemandedBits handling.
Differential Revision: https://reviews.llvm.org/D52980
llvm-svn: 344019
Simple types are a superset of what all in tree targets in LLVM could possibly have a legal type. This means the behavior of using isSimple to check for a supported type for X86 could change over time. For example, this could would change if a v256i1 type was added to MVT in the future.
llvm-svn: 343995
This patch implements a pass that optimizes condition branches on x86 by
taking advantage of the three-way conditional code generated by compare
instructions.
Currently, it tries to hoisting EQ and NE conditional branch to a dominant
conditional branch condition where the same EQ/NE conditional code is
computed. An example:
bb_0:
cmp %0, 19
jg bb_1
jmp bb_2
bb_1:
cmp %0, 40
jg bb_3
jmp bb_4
bb_4:
cmp %0, 20
je bb_5
jmp bb_6
Here we could combine the two compares in bb_0 and bb_4 and have the
following code:
bb_0:
cmp %0, 20
jg bb_1
jl bb_2
jmp bb_5
bb_1:
cmp %0, 40
jg bb_3
jmp bb_6
For the case of %0 == 20 (bb_5), we eliminate two jumps, and the control height
for bb_6 is also reduced. bb_4 is gone after the optimization.
This optimization is motivated by the branch pattern generated by the switch
lowering: we always have pivot-1 compare for the inner nodes and we do a pivot
compare again the leaf (like above pattern).
This pass currently is enabled on Intel's Sandybridge and later arches. Some
reviewers pointed out that on some arches (like AMD Jaguar), this pass may
increase branch density to the point where it hurts the performance of the
branch predictor.
Differential Revision: https://reviews.llvm.org/D46662
llvm-svn: 343993
Some necessary yak shaving before lowering *_EXTEND_VECTOR_INREG 256-bit vectors on AVX1 targets as suggested by D52964.
Differential Revision: https://reviews.llvm.org/D52970
llvm-svn: 343991
The instructions are complicated, so this code will
probably never be very obvious, but hopefully this
makes it better.
As shown in PR39195:
https://bugs.llvm.org/show_bug.cgi?id=39195
...we need to improve the matching to not miss cases
where we're h-opping on 1 source vector, and that
should be a small patch after this rearranging.
llvm-svn: 343989
Support G_UDIV/G_UREM/G_SREM. The instruction selection
code is taken from FastISel with only minor tweaks to adapt
for GlobalISel.
Differential Revision: https://reviews.llvm.org/D49781
llvm-svn: 343966
Prevents missing other simplifications that may occur deep in the operand chain where CommitTargetLoweringOpt won't add the PMULDQ back to the worklist itself
llvm-svn: 343922
Attempt to simplify PSHUFB masks (even non-constant ones) - we should probably be able to simplify other variable shuffles as well as the need arises.
llvm-svn: 343919
This rebases and recommits r343520. hwasan should be fixed now and this
shouldn't break the tests anymore.
Spill/reload instructions are artificially generated by the compiler and
have no relation to the original source code. So the best thing to do is
not attach any debug location to them (instead of just taking the next
debug location we find on following instructions).
Differential Revision: https://reviews.llvm.org/D52125
llvm-svn: 343895
rL343853 didn't limit the number of subinputs, but we don't currently support faux shuffles with more than 2 total inputs, so put a limiter in place until this is fixed.
Found by Artem Dergachev.
llvm-svn: 343891
The comments in this code say we were trying to avoid 16-bit immediates, but if the immediate fits in 8-bits this isn't an issue. This avoids creating a zero extend that probably won't go away.
The movmskb related changes are interesting. The movmskb instruction writes a 32-bit result, but fills the upper bits with 0. So the zero_extend we were previously emitting was free, but we turned a -1 immediate that would fit in 8-bits into a 32-bit immediate so it was still bad.
llvm-svn: 343871
Currently we hardcode instructions with ReadAfterLd if the register operands don't need to be available until the folded load has completed. This doesn't take into account the different load latencies of different memory operands (PR36957).
This patch adds a ReadAfterFold def into X86FoldableSchedWrite to replace ReadAfterLd, allowing us to specify the load latency at a scheduler class level.
I've added ReadAfterVec*Ld classes that match the XMM/Scl, XMM and YMM/ZMM WriteVecLoad classes that we currently use, we can tweak these values in future patches once this infrastructure is in place.
Differential Revision: https://reviews.llvm.org/D52886
llvm-svn: 343868
Decode subvector shuffles from INSERT_SUBVECTOR(SRC0, SHUFFLE(EXTRACT_SUBVECTOR(SRC1))
This was found necessary while investigating PR39161
llvm-svn: 343853
Finally all targets are enabling multiple regalloc hints, so the hook to
disable this can now be removed.
NFC.
Review: Simon Pilgrim
https://reviews.llvm.org/D52316
llvm-svn: 343851
Previously we replaced the chain use ourself and return the data result. LegalizeVectorOps then detected that we'd done this and assumed the chain had already been handled.
This commit instead returns a MERGE_VALUES node with two results joined from nodes. This allows LegalizeVectorOps to do all the replacements for us without any special casing. The MERGE_VALUES will be removed by DAG combine.
llvm-svn: 343817
This can happen if assembling a reference to _GLOBAL_OFFSET_TABLE_.
While it doesn't make sense to try to assemble that for COFF,
the fact that we previously used llvm_unreachable meant that the code
had undefined behaviour if something tried to assemble that.
The configure script of libgmp would try to assemble such a snippet
(which should signal a failure). If llvm is built without assertions,
the undefined behaviour meant a (near) infinite loop.
Differential Revision: https://reviews.llvm.org/D52903
llvm-svn: 343811
The additional patterns needed for this aren't overwhelming and introducing extra bitcasts during lowering limits our ability to do computeNumSignBits. Not that I have a good example of that for select. I'm just becoming increasingly grumpy about promotion of AND/OR/XOR. SELECT was just a lot easier to fix.
llvm-svn: 343723
This patch adds a 'WriteCopy' [WriteLoad, WriteStore] schedule sequence instead to better model the behaviour
Found by @andreadb during llvm-mca testing on btver2 which was crashing on "zero uop" WriteRMW only instructions
llvm-svn: 343708
Fix use of SSE1 registers for f32 ops in no-x87 mode.
Notably, allow use of SSE instructions for f32 operations in 64-bit
mode (but not 32-bit which is disallowed by callign convention).
Also avoid translating memset/memcopy/memmove into SSE registers
without X87 for 32-bit mode.
This fixes PR38738.
Reviewers: nickdesaulniers, craig.topper
Subscribers: hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D52555
llvm-svn: 343689
I was expecting this to be a nfc but Silvermont seems to be setup a little differently:
// A folded store needs a cycle on MEC_RSV for the store data, but it does not need an extra port cycle to recompute the address.
def : WriteRes<WriteRMW, [SLM_MEC_RSV]>;
So moving from WriteStore to WriteRMW reduces predicted port pressure, confirmed by @craig.topper that this is correct.
Differential Revision: https://reviews.llvm.org/D52740
llvm-svn: 343670
The 0x63 opcodes in 64-bit mode have a fixed source size of 32-bits, but the destination size is controlled by REX.W and the 0x66 opsize prefix. This instruction is normally used with a REX.W prefix which provides desired behavior. The other encodings are interpretted as valid by the processor, but aren't useful.
This patch makes us recognize them for the disassembler to match objdump.
llvm-svn: 343614
Add the .cv_fpo_stackalign directive so that we can define $T0, or the
VFRAME virtual register, with it. This was overlooked in the initial
implementation because unlike MSVC, we push CSRs before allocating stack
space, so this value is only needed to describe local variable
locations. Variables that the compiler now addresses via ESP are instead
described as being stored at offsets from VFRAME, which for us is ESP
after alignment in the prologue.
This adds tests that show that we use the VFRAME register properly in
our S_DEFRANGE records, and that we emit the correct FPO data to define
it.
Fixes PR38857
llvm-svn: 343603
Consistently try to use APFloat::toString for floating point constant comments to get rid of differences between Constant / ConstantDataSequential values - it should help stop some of the linux-windows buildbot failures matching NaN/INF etc. as well.
Differential Revision: https://reviews.llvm.org/D52702
llvm-svn: 343562
Summary:
Before this change, LLVM would always describe locals on the stack as
being relative to some specific register, RSP, ESP, EBP, ESI, etc.
Variables in stack memory are pretty common, so there is a special
S_DEFRANGE_FRAMEPOINTER_REL symbol for them. This change uses it to
reduce the size of our debug info.
On top of the size savings, there are cases on 32-bit x86 where local
variables are addressed from ESP, but ESP changes across the function.
Unlike in DWARF, there is no FPO data to describe the stack adjustments
made to push arguments onto the stack and pop them off after the call,
which makes it hard for the debugger to find the local variables in
frames further up the stack.
To handle this, CodeView has a special VFRAME register, which
corresponds to the $T0 variable set by our FPO data in 32-bit. Offsets
to local variables are instead relative to this value.
This is part of PR38857.
Reviewers: hans, zturner, javed.absar
Subscribers: aprantl, hiraditya, JDevlieghere, llvm-commits
Differential Revision: https://reviews.llvm.org/D52217
llvm-svn: 343543
This includes a fix to prevent i16 compares with i32/i64 ands from being shrunk if bit 15 of the and is set and the sign bit is used.
Original commit message:
Currently we skip looking through truncates if the sign flag is used. But that's overly restrictive.
It's safe to look through the truncate as long as we ensure one of the 3 things when we shrink. Either the MSB of the mask at the shrunken size isn't set. If the mask bit is set then either the shrunk size needs to be equal to the compare size or the sign
There are still missed opportunities to shrink a load and fold it in here. This will be fixed in a future patch.
llvm-svn: 343539
Spill/reload instructions are artificially generated by the compiler and
have no relation to the original source code. So the best thing to do is
not attach any debug location to them (instead of just taking the next
debug location we find on following instructions).
Differential Revision: https://reviews.llvm.org/D52125
llvm-svn: 343520
There's a subtle bug in the handling of truncate from i32/i64 to i32 without minsize.
I'll be adding more test cases and trying to find a fix.
llvm-svn: 343516
This patch adds load folding support to the test shrinking code. This was noticed missing in the review for D52669
Differential Revision: https://reviews.llvm.org/D52699
llvm-svn: 343499
Currently we skip looking through truncates if the sign flag is used. But that's overly restrictive.
It's safe to look through the truncate as long as we ensure one of the 3 things when we shrink. Either the MSB of the mask at the shrunken size isn't set. If the mask bit is set then either the shrunk size needs to be equal to the compare size or the sign flag needs to be unused.
There are still missed opportunities to shrink a load and fold it in here. This will be fixed in a future patch.
Differential Revision: https://reviews.llvm.org/D52669
llvm-svn: 343498
This patch adds another variant class to identify zero-idiom VPERM2F128rr
instructions.
On Jaguar, a VPERM wih bit 3 and 7 of the mask set, is a zero-idiom.
Differential Revision: https://reviews.llvm.org/D52663
llvm-svn: 343452
Summary:
While looking at PR35606, I found out that the scheduling info is incorrect.
One can check that it's really a P5+P6 and not a 2*P56 with:
echo -e 'vzeroall\nvandps %xmm1, %xmm2, %xmm3' | ./bin/llvm-exegesis -mode=uops -snippets-file=-
(vandps executes on P5 only)
Reviewers: craig.topper, RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D52541
llvm-svn: 343447
We can only copy between a k-register and a GR32/GR64 register.
This patch detects that the copy will be illegal and prevents the domain reassignment from happening for that closure.
This probably isn't the best fix, and we should probably figure out how to handle this correctly.
Fixes PR38803.
llvm-svn: 343443
There's a conditional report_fatal_error just above this llvm_unreachable. The optimizer when seeing the unreachable removes the conditional and just makes any other error trigger the existing report_fatal_error.
llvm-svn: 343428
Summary:
This function turns (X >> C1) & C2 into a BMI BEXTR or TBM BEXTRI instruction. For BMI BEXTR we have to materialize an immediate into a register to feed to the BEXTR instruction.
The BMI BEXTR instruction is 2 uops on Intel CPUs. It looks like on SKL its one port 0/6 uop and one port 1/5 uop. Despite what Agner's tables say. I know one of the uops is a regular shift uop so it would have to go through the port 0/6 shifter unit. So that's the same or worse execution wise than the shift+and which is one 0/6 uop and one 0/1/5/6 uop. The move immediate into register is an additional 0/1/5/6 uop.
For now I've limited this transform to AMD CPUs which have a single uop BEXTR. If may also might make sense if we can fold a load or if the and immediate is larger than 32-bits and can't be encoded as a sign extended 32-bit value or if LICM or CSE can hoist the move immediate and share it. But we'd need to look more carefully at that. In the regression I looked at it doesn't look load folding or large immediates were occurring so the regression isn't caused by the loss of those. So we could try to be smarter here if we find a compelling case.
Reviewers: RKSimon, spatel, lebedev.ri, andreadb
Reviewed By: RKSimon
Subscribers: llvm-commits, andreadb, RKSimon
Differential Revision: https://reviews.llvm.org/D52570
llvm-svn: 343399
By removing demanded target shuffles that simplify to zero/undef/identity before simplifying its inputs we improve chances of further simplification, as only the immediate parent user of the combined is added back to the work list - this still doesn't help us if its passed through other ops though (bitcasts....).
llvm-svn: 343390
The shift amount might have peeked through a extract_subvector, altering the number of vector elements in the 'Amt' variable - so we were incorrectly calculating the ratio when peeking through bitcasts, resulting in incorrectly detecting splats.
llvm-svn: 343373
Similar to the existing ISD::SRL constant vector shifts from D49562, this patch adds ISD::SRA support with ISD::MULHS.
As we're dealing with signed values, we have to handle shift by zero and shift by one special cases, so XOP+AVX2/AVX512 splitting/extension is still a better solution - really we should still use ISD::MULHS if one of the special cases are used but for now I've just left a TODO and filtered by isKnownNeverZero.
Differential Revision: https://reviews.llvm.org/D52171
llvm-svn: 343093
This removes an int->fp bitcast between the surrounding code and the movmsk. I had already added a hack to combineMOVMSK to try to look through this bitcast to improve the SimplifyDemandedBits there.
But I found an additional issue where the bitcast was preventing combineMOVMSK from being called again after earlier nodes in the DAG are optimized. The bitcast gets revisted, but not the user of the bitcast. By using integer types throughout, the bitcast doesn't get in the way.
llvm-svn: 343046
This is the final (I hope!) problem pattern mentioned in PR37749:
https://bugs.llvm.org/show_bug.cgi?id=37749
We are trying to avoid an AVX1 sinkhole caused by having 256-bit bitwise logic ops but no other 256-bit integer ops.
We've already solved the simple logic ops, but 'andn' is an x86 special. I looked at alternative solutions like
extending the generic DAG combine or trying to wait until the ANDNP node is created, but those are bigger patches
that can over-reach. Ie, splitting to 128-bit does not look like a win in most cases with >1 256-bit op.
The pattern matching is cluttered with bitcasts because of our i64 element canonicalization. For the affected test,
we have this vector-type-legalized sequence:
t29: v8i32 = concat_vectors t27, t28
t30: v4i64 = bitcast t29
t18: v8i32 = BUILD_VECTOR Constant:i32<-1>, Constant:i32<-1>, ...
t31: v4i64 = bitcast t18
t32: v4i64 = xor t30, t31
t9: v8i32 = BUILD_VECTOR Constant:i32<255>, Constant:i32<255>, ...
t34: v4i64 = bitcast t9
t35: v4i64 = and t32, t34
t36: v8i32 = bitcast t35
t37: v4i32 = extract_subvector t36, Constant:i64<0>
t38: v4i32 = extract_subvector t36, Constant:i64<4>
Differential Revision: https://reviews.llvm.org/D52318
llvm-svn: 343008
As suggested by Craig Topper - I'm going to look at cleaning up the RMW sequences instead.
The uops are slightly different to the register variant, so requires a +1uop tweak
llvm-svn: 342969
We're missing quite a bit of data for these instruction, removing the overrides makes this obvious - inconsistent reg/mem variants is a concern as well.
Also, we have Divider resources (HWDivider etc.) but they aren't actually used consistently.
llvm-svn: 342904
Split WriteIMul by size and also by IMUL multiply-by-imm and multiply-by-reg cases.
This removes all the scheduler overrides for gpr multiplies and stops WriteMULH being ignored for BMI2 MULX instructions.
llvm-svn: 342892
Variable Shifts/Rotates using the CL register have different behaviours to the immediate instructions - split accordingly to help remove yet more repeated overrides from the schedule models.
llvm-svn: 342852
Confirmed with Craig Topper - fix a typo that was missing a Port4 uop for ROR*mCL instructions on some Intel models.
Yet another step on the scheduler model cleanup marathon......
llvm-svn: 342846
This is an alternative to https://reviews.llvm.org/D37896. We can't decompose
multiplies generically without a target hook to tell us when it's profitable.
ARM and AArch64 may be able to remove some existing code that overlaps with
this transform.
This extends D52195 and may resolve PR34474:
https://bugs.llvm.org/show_bug.cgi?id=34474
(still an open question about transforming legal vector multiplies, but we
could open another bug report for those)
llvm-svn: 342844
The SandyBridge model was missing schedule values for the RCL/RCR values - instead using the (incredibly optimistic) WriteShift (now WriteRotate) defaults.
I've added overrides with more realistic (slow) values, based on a mixture of Agner/instlatx64 numbers and what later Intel models do as well.
This is necessary to allow WriteRotate to be updated to remove other rotate overrides.
It'd probably be a good idea to investigate a WriteRotateCarry class at some point but its not high priority given the unusualness of these instructions.
llvm-svn: 342842
Despite being rotates, these more modern instructions avoid many of the quirks of the regular x86 rotate instructions and consistently have a schedule closer to shifts.
llvm-svn: 342839
NFCI for now, but it should make it easier to remove a lot of unnecessary overrides in a future commit.
Now that funnel shift intrinsics are coming online we need to get this cleaned up to make vectorization costs from scalar rotate patterns more straightforward.
llvm-svn: 342837
Our lowering that tries to avoid this sign extend can be defeated by the DAG combine folding it with a truncate.
The pattern needs to extend to an v8i32 then truncate back down to v8i16.
llvm-svn: 342830
Summary: Similar to D51893 which was for memcpy
Reviewers: efriedma
Reviewed By: efriedma
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D52063
llvm-svn: 342796
We don't have a vXi8 shift left so we need to bitcast to a vXi16 vector to perform the shift. If we let lowering legalize the vXi8 shift we get an extra and that we don't need and fail to remove.
llvm-svn: 342795
Previously we used SUBREG_TO_REG+MOV32ri. But regular isel was changed recently to use the MOV32ri64 pseudo. Fast isel now does the same.
llvm-svn: 342788
This patch introduces a SchedWriteVariant to describe zero-idiom VXORP(S|D)Yrr
and VANDNP(S|D)Yrr.
This is a follow-up of r342555.
On Jaguar, a VXORPSYrr is 2 macro opcodes. Only one opcode is eliminated at
register-renaming stage. The other opcode has to be executed to set the upper
half of the destination YMM.
Same for VANDNP(S|D)Yrr.
Differential Revision: https://reviews.llvm.org/D52347
llvm-svn: 342728
x86 had 2 versions of peekThroughBitcast. DAGCombiner had 1. Plus, it had a 1-off implementation for the one-use variant.
Move the x86 versions of the code to SelectionDAG, so we don't have different copies of the code.
No functional change intended.
I'm putting this next to isBitwiseNot() because I am planning to use it in there. Another option is next to the
helpers in the ISD namespace (eg, ISD::isConstantSplatVector()). But if there's no good reason for those to be
there, I'd prefer to pull other helpers over to SelectionDAG in follow-up steps.
Differential Revision: https://reviews.llvm.org/D52285
llvm-svn: 342669
Enable enableMultipleCopyHints() on X86.
Original Patch by @jonpa:
While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling.
Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates.
Differential Revision: https://reviews.llvm.org/D38128
llvm-svn: 342578
As the code comments suggest, these are about splitting, and they
are not necessarily limited to lowering, so that misled me.
There's nothing that's actually x86-specific in these either, so
they might be better placed in a common header so any target can
use them.
llvm-svn: 342575
This patch adds an initial x86 SimplifyDemandedVectorEltsForTargetNode implementation to handle target shuffles.
Currently the patch only decodes a target shuffle, calls SimplifyDemandedVectorElts on its input operands and removes any shuffle that reduces to undef/zero/identity.
Future work will need to integrate this with combineX86ShufflesRecursively, add support for other x86 ops, etc.
NOTE: There is a minor regression that appears to be affecting further (extractelement?) combines which I haven't been able to solve yet - possibly something to do with how nodes are added to the worklist after simplification.
Differential Revision: https://reviews.llvm.org/D52140
llvm-svn: 342564
This patch adds the ability for processor models to describe dependency breaking
instructions.
Different processors may specify a different set of dependency-breaking
instructions.
That means, we cannot assume that all processors of the same target would use
the same rules to classify dependency breaking instructions.
The main goal of this patch is to provide the means to describe dependency
breaking instructions directly via tablegen, and have the following
TargetSubtargetInfo hooks redefined in overrides by tabegen'd
XXXGenSubtargetInfo classes (here, XXX is a Target name).
```
virtual bool isZeroIdiom(const MachineInstr *MI, APInt &Mask) const {
return false;
}
virtual bool isDependencyBreaking(const MachineInstr *MI, APInt &Mask) const {
return isZeroIdiom(MI);
}
```
An instruction MI is a dependency-breaking instruction if a call to method
isDependencyBreaking(MI) on the STI (TargetSubtargetInfo object) evaluates to
true. Similarly, an instruction MI is a special case of zero-idiom dependency
breaking instruction if a call to STI.isZeroIdiom(MI) returns true.
The extra APInt is used for those targets that may want to select which machine
operands have their dependency broken (see comments in code).
Note that by default, subtargets don't know about the existence of
dependency-breaking. In the absence of external information, those method calls
would always return false.
A new tablegen class named STIPredicate has been added by this patch to let
processor models classify instructions that have properties in common. The idea
is that, a MCInstrPredicate definition can be used to "generate" an instruction
equivalence class, with the idea that instructions of a same class all have a
property in common.
STIPredicate definitions are essentially a collection of instruction equivalence
classes.
Also, different processor models can specify a different variant of the same
STIPredicate with different rules (i.e. predicates) to classify instructions.
Tablegen backends (in this particular case, the SubtargetEmitter) will be able
to process STIPredicate definitions, and automatically generate functions in
XXXGenSubtargetInfo.
This patch introduces two special kind of STIPredicate classes named
IsZeroIdiomFunction and IsDepBreakingFunction in tablegen. It also adds a
definition for those in the BtVer2 scheduling model only.
This patch supersedes the one committed at r338372 (phabricator review: D49310).
The main advantages are:
- We can describe subtarget predicates via tablegen using STIPredicates.
- We can describe zero-idioms / dep-breaking instructions directly via
tablegen in the scheduling models.
In future, the STIPredicates framework can be used for solving other problems.
Examples of future developments are:
- Teach how to identify optimizable register-register moves
- Teach how to identify slow LEA instructions (each subtarget defining its own
concept of "slow" LEA).
- Teach how to identify instructions that have undocumented false dependencies
on the output registers on some processors only.
It is also (in my opinion) an elegant way to expose knowledge to both external
tools like llvm-mca, and codegen passes.
For example, machine schedulers in LLVM could reuse that information when
internally constructing the data dependency graph for a code region.
This new design feature is also an "opt-in" feature. Processor models don't have
to use the new STIPredicates. It has all been designed to be as unintrusive as
possible.
Differential Revision: https://reviews.llvm.org/D52174
llvm-svn: 342555
This is an alternative to D37896. I don't see a way to decompose multiplies
generically without a target hook to tell us when it's profitable.
ARM and AArch64 may be able to remove some duplicate code that overlaps with
this transform.
As a first step, we're only getting the most clear wins on the vector examples
requested in PR34474:
https://bugs.llvm.org/show_bug.cgi?id=34474
As noted in the code comment, it's likely that the x86 constraints are tighter
than necessary, but it may not always be a win to replace a pmullw/pmulld.
Differential Revision: https://reviews.llvm.org/D52195
llvm-svn: 342554
The 0x800 bit in @feat.00 needs to be set in order to make LLD pick up
the .gfid$y table. I believe this is fine to set even if we don't emit
the instrumentation.
We haven't emitted @feat.00 on 64-bit before. I see that MSVC does emit
it, but I'm not entirely sure what the default value should be. I went
with zero since that seems as safe as not emitting the symbol in the
first place.
Differential Revision: https://reviews.llvm.org/D52235
llvm-svn: 342532
Summary:
The IR reference for the `byval` attribute states:
```
This indicates that the pointer parameter should really be passed by value
to the function. The attribute implies that a hidden copy of the pointee is
made between the caller and the callee, so the callee is unable to modify
the value in the caller. This attribute is only valid on LLVM pointer arguments.
```
However, on Win64, this attribute is unimplemented and the raw pointer is
passed to the callee instead. This is problematic, because frontend authors
relying on the implicit hidden copy (as happens for every other calling
convention) will see the passed value silently (if mutable memory) or
loudly (by means of a crash) modified because the callee treats the
location as scratch memory space it is allowed to mutate.
At this point, it's worth taking a step back to understand the context.
In most calling conventions, aggregates that are too large to be passed
in registers, instead get *copied* to the stack at a fixed (computable
from the signature) offset of the stack pointer. At the LLVM, we hide
this hidden copy behind the byval attribute. The caller passes a pointer
to the desired data and the callee receives a pointer, but these pointers
are not the same. In particular, the pointer that the callee receives
points to temporary stack memory allocated as part of the call lowering.
In most calling conventions, this pointer is never realized in registers
or memory. The temporary memory is simply defined by an implicit
offset from the stack pointer at function entry.
Win64, uniquely, works differently. The structure is still passed in
memory, but instead of being stored at an implicit memory offset, the
caller computes a pointer to the temporary memory and passes it to
the callee as a regular pointer (taking up a register, or if all
registers are taken up, an additional stack slot). Presumably, this
was done to allow eliding the copy when passing aggregates through
several functions on the stack.
This explains why ignoring the `byval` attribute mostly works on Win64.
The argument simply gets passed as a pointer and as long as we're ok
with the callee trampling all over that memory, there are no ill effects.
However, it does contradict the documentation of the `byval` attribute
which specifies that there is to be an implicit copy.
Frontends can of course work around this by never emitting the `byval`
attribute for Win64 and creating `alloca`s for the requisite temporary
stack slots (and that does appear to be what frontends are doing).
However, the presence of the `byval` attribute is not a trap for
frontend authors, since it seems to work, but silently modifies the
passed memory contrary to documentation.
I see two solutions:
- Disallow the `byval` attribute in the verifier if using the Win64
calling convention.
- Make it work by simply emitting a temporary stack copy as we would
with any other calling convention (frontends can of course always
not use the attribute if they want to elide the copy).
This patch implements the second option (make it work), though I would
be fine with the first also.
Ref: https://github.com/JuliaLang/julia/issues/28338
Reviewers: rnk
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51842
llvm-svn: 342402
For constant non-uniform cases we'll never introduce more and/andn/or selects than already occur in generic pre-SSE41 ISD::SRL lowering.
llvm-svn: 342352
https://bugs.llvm.org/show_bug.cgi?id=38949
It's not clear to me that we even need a one-use check in this fold.
Ie, 2 independent loads might be better than a load+dependent shuffle.
Note that the existing re-use tests are not affected. We actually do form a
broadcast node in those tests now because there's no extra use of the
insert_subvector node in those cases. But something later in isel pattern
matching decides that it is not worth using a broadcast for the full load in
those tests:
Legalized selection DAG: %bb.0 'test_broadcast_2f64_4f64_reuse:'
t7: v2f64,ch = load<(load 16 from %ir.p0)> t0, t2, undef:i64
t4: i64,ch = CopyFromReg t0, Register:i64 %1
t10: ch = store<(store 16 into %ir.p1)> t7:1, t7, t4, undef:i64
t18: v4f64 = insert_subvector undef:v4f64, t7, Constant:i64<0>
t20: v4f64 = insert_subvector t18, t7, Constant:i64<2>
Becomes:
t7: v2f64,ch = load<(load 16 from %ir.p0)> t0, t2, undef:i64
t4: i64,ch = CopyFromReg t0, Register:i64 %1
t10: ch = store<(store 16 into %ir.p1)> t7:1, t7, t4, undef:i64
t21: v4f64 = X86ISD::SUBV_BROADCAST t7
ISEL: Starting selection on root node: t21: v4f64 = X86ISD::SUBV_BROADCAST t7
...
Created node: t27: v4f64 = INSERT_SUBREG IMPLICIT_DEF:v4f64, t7, TargetConstant:i32<7>
Morphed node: t21: v4f64 = VINSERTF128rr t27, t7, TargetConstant:i8<1>
llvm-svn: 342347
Summary: This unfortunately adds a move, but isn't that better than going to the int domain and back?
Reviewers: RKSimon
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D52134
llvm-svn: 342327
Summary:
MOVMSK only care about the sign bit so we don't need the setcc to fill the whole element with 0s/1s. We can just shift the bit we're looking for into the sign bit. This saves a constant pool load.
Inspired by PR38840.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: lebedev.ri, llvm-commits
Differential Revision: https://reviews.llvm.org/D52121
llvm-svn: 342326
Attempt to lower a shuffle as an unpack of elements from two inputs followed by a single-input (wider) permutation.
As long as the permutation is wider this is a win - there may be some circumstances where same size permutations would also be useful but I've left that for future work.
Differential Revision: https://reviews.llvm.org/D52043
llvm-svn: 342257
When replacing a named register input to the appropriately sized
sub/super-register. In the case of a 64-bit value being assigned to a
register in 32-bit mode, match GCC's assignment.
Reviewers: eli.friedman, craig.topper
Subscribers: nickdesaulniers, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D51502
llvm-svn: 342175
Summary:
Previously we type legalized v2i32 div/rem by promoting to v2i64. But we don't support div/rem of vectors so op legalization would then scalarize it using i64 scalar ops since it doesn't know about the original promotion. 64-bit scalar divides on Intel hardware are known to be slow and in 32-bit mode they require a libcall.
This patch switches type legalization to do the scalarizing itself using i32.
It looks like the division by power of 2 optimization is still kicking in and leaving the code as a vector. The division by other constant optimization doesn't kick in pre type legalization since it ignores illegal types. And previously, after type legalization we scalarized the v2i64 since we don't have v2i64 MULHS/MULHU support.
Another option might be to widen v2i32 to v4i32 so we could do division by constant optimizations, but we'd have to be careful to only do that for constant divisors or we risk scalaring to 4 scalar divides.
Reviewers: RKSimon, spatel
Reviewed By: spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51325
llvm-svn: 342114
There's no advantage to this instruction unless you need to avoid touching other flag bits. It's encoding is longer, it can't fold an immediate, it doesn't write all the flags.
I don't think gcc will generate this instruction either.
Fixes PR38852.
Differential Revision: https://reviews.llvm.org/D51754
llvm-svn: 342059
Summary:
In GNUX23, is64BitMode returns true, but pointers are 32-bits. So we shouldn't copy pointer values into RSI/RDI since the widths don't match.
Fixes PR38865 despite what the title says. I think the llvm_unreachable in the copyPhysReg code tricked the optimizer and made the fatal error trigger.
Reviewers: rnk, efriedma, MatzeB, echristo
Reviewed By: efriedma
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51893
llvm-svn: 342015
In r337348, I changed lowering to prefer X86ISD::UNPCKL/UNPCKH opcodes over MOVLHPS/MOVHLPS for v2f64 {0,0} and {1,1} shuffles when we have SSE2. This enabled the removal of a bunch of weirdly bitcasted isel patterns in r337349. To avoid changing the tests I placed a gross hack in isel to still emit movhlps instructions for fake unary unpckh nodes. A similar hack was not needed for unpckl and movlhps because we do execution domain switching for those. But unpckh and movhlps have swapped operand order.
This patch removes the hack.
This is a code size increase since unpckhpd requires a 0x66 prefix and movhlps does not. But if that's a big concern we should be using movhlps for all unpckhpd opcodes and let commuteInstruction turnit into unpckhpd when its an advantage.
Differential Revision: https://reviews.llvm.org/D49499
llvm-svn: 341973
GNUX32 uses 32-bit pointers despite is64BitMode being true. So we should use EAX to return the value.
Fixes ones of the failures from PR38865.
Differential Revision: https://reviews.llvm.org/D51940
llvm-svn: 341972
MOVMSKPS and MOVMSKPD both take FP types, but likely the operations before it are on integer types with just a int->fp bitcast between them. If the bitcast isn't used by anything else and doesn't change the element width we can look through it to simplify the integer ops.
llvm-svn: 341915
I'm having a hard time finding a test case for this, but we should be consistent here. The fact that we canonicalize all zeros and all ones constants to vXi32 and all other constants to loads makes this hard to hit the easy DAG combine infinite loop we get for some of the other types.
llvm-svn: 341859
We have isel patterns for v4i32/v4f64 that artificially widen to v8i32/v8f64 so just use that.
If x86-experimental-vector-widening-legalization is enabled, we don't need any custom legalization and can just return. I've modified the test RUN lines to cover this case.
llvm-svn: 341765
Summary:
This patch allows vectors with a power of 2 number of elements and i8/i16 element type to select paddus/psubus instructions. ReplaceNodeResults has been updated to custom widen these operations up to 128 bits like we already do for PAVG.
Another step towards fixing PR38691
Reviewers: RKSimon, spatel
Reviewed By: RKSimon, spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51818
llvm-svn: 341753
The generic type legalizer will scalarize vXi1 instructions getting rid of the vector entirely. Creating wider vector instructions is just going to prevent that.
llvm-svn: 341705
The type legalizer will try to scalarize this and fail.
It looks like there's some other v1iX oddities out there too since we still generated some vector instructions.
llvm-svn: 341704
Similar to what was recently done for addcarry/subborrow and has been done for rdrand/rdseed for a while. It's better to use two results and an explicit store in IR when the store isn't part of the semantics of the instruction. This allows store->load forwarding to happen in the middle end. Or the store to be removed if its never loaded.
Differential Revision: https://reviews.llvm.org/D51803
llvm-svn: 341698
We should represent the store directly in IR instead. This gives the middle end a chance to remove it if it can see a load from the same address.
Differential Revision: https://reviews.llvm.org/D51769
llvm-svn: 341677
Previously we only handled loads in operand 0, but nothing guarantees the load will be operand 0 for commutable operations.
Differential Revision: https://reviews.llvm.org/D51768
llvm-svn: 341675
ADC is commutable and the load could be in either operand, but we were only checking operand 0.
Ideally we'd mark X86adc_flag as commutable and tablegen would automatically do this, but the EFLAGS register mention is preventing it.
llvm-svn: 341606
The peephole pass likely gets this normally, but we should be doing it during isel.
Ideally we'd just make the X86adc_flag pattern SDNPCommutable, but the tablegen doesn't handle that when one of the operands is a register reference.
llvm-svn: 341596
This basically reverts a change made in r336217, but improves the text of the error message for not allowing IP-relative addressing in 32-bit mode.
Fixes PR38826.
Patch by Iain Sandoe.
llvm-svn: 341512
This removes the FrameAccess struct that was added to the interface
in D51537, since the PseudoValue from the MachineMemoryOperand
can be safely casted to a FixedStackPseudoSourceValue.
Reviewers: MatzeB, thegameg, javed.absar
Reviewed By: thegameg
Differential Revision: https://reviews.llvm.org/D51617
llvm-svn: 341454
On Windows, if shouldAssumeDSOLocal returns false, it's either a
dllimport reference, or a reference that we should treat as non-local
and create a stub for.
Clean up AArch64Subtarget::ClassifyGlobalReference a little while
touching the flag handling relating to dllimport.
Differential Revision: https://reviews.llvm.org/D51590
llvm-svn: 341402
Load Hardening.
Wires up the existing pass to work with a proper IR attribute rather
than just a hidden/internal flag. The internal flag continues to work
for now, but I'll likely remove it soon.
Most of the churn here is adding the IR attribute. I talked about this
Kristof Beyls and he seemed at least initially OK with this direction.
The idea of using a full attribute here is that we *do* expect at least
some forms of this for other architectures. There isn't anything
*inherently* x86-specific about this technique, just that we only have
an implementation for x86 at the moment.
While we could potentially expose this as a Clang-level attribute as
well, that seems like a good question to defer for the moment as it
isn't 100% clear whether that or some other programmer interface (or
both?) would be best. We'll defer the programmer interface side of this
for now, but at least get to the point where the feature can be enabled
without relying on implementation details.
This also allows us to do something that was really hard before: we can
enable *just* the indirect call retpolines when using SLH. For x86, we
don't have any other way to mitigate indirect calls. Other architectures
may take a different approach of course, and none of this is surfaced to
user-level flags.
Differential Revision: https://reviews.llvm.org/D51157
llvm-svn: 341363
implementing the proposed mitigation technique described in the original
design document.
The idea is to check after calls that the return address used to arrive
at that location is in fact the correct address. In the event of
a mis-predicted return which reaches a *valid* return but not the
*correct* return, this will detect the mismatch much like it would
a mispredicted conditional branch.
This is the last published attack vector that I am aware of in the
Spectre v1 space which is not mitigated by SLH+retpolines. However,
don't read *too* much into that: this is an area of ongoing research
where we expect more issues to be discovered in the future, and it also
makes no attempt to mitigate Spectre v4. Still, this is an important
completeness bar for SLH.
The change here is of course delightfully simple. It was predicated on
cutting support for post-instruction symbols into LLVM which was not at
all simple. Many thanks to Hal Finkel, Reid Kleckner, and Justin Bogner
who helped me figure out how to do a bunch of the complex changes
involved there.
Differential Revision: https://reviews.llvm.org/D50837
llvm-svn: 341358
retpolines.
This implements the core design of tracing the intended target into the
target, checking it, and using that to update the predicate state. It
takes advantage of a few interesting aspects of SLH to make it a bit
easier to implement:
- We already split critical edges with conditional branches, so we can
assume those are gone.
- We already unfolded any memory access in the indirect branch
instruction itself.
I've left hard errors in place to catch if any of these somewhat subtle
invariants get violated.
There is some code that I can factor out and share with D50837 when it
lands, but I didn't want to couple landing the two patches, so I'll do
that in a follow-up cleanup commit if alright.
Factoring out the code to handle different scenarios of materializing an
address remains frustratingly hard. In a bunch of cases you want to fold
one of the cases into an immediate operand of some other instruction,
and you also have both symbols and basic blocks being used which require
different methods on the MI builder (and different operand kinds).
Still, I'll take a stab at sharing at least some of this code in
a follow-up if I can figure out how.
Differential Revision: https://reviews.llvm.org/D51083
llvm-svn: 341356
A ReadAdvance was incorrectly added to the SchedReadWrite list associated with
the following SSE instructions:
sqrtss
sqrtsd
rsqrtss
rcpss
As a consequence, a wrong operand latency was computed for the register operand
used as the base address of the folded load operand.
This patch removes the wrong ReadAdvance, and updates the llvm-mca test cases.
There is still a problem with correctly modeling partial register writes on XMM
registers This other problem is currently tracked here:
https://bugs.llvm.org/show_bug.cgi?id=38813
Differential Revision: https://reviews.llvm.org/D51542
llvm-svn: 341326
Also adjust some of dsymutil's headers to put the header guards at the top,
otherwise the compiler will not recognize them as header guards.
llvm-svn: 341323
For instructions that spill/fill to and from multiple frame-indices
in a single instruction, hasStoreToStackSlot and hasLoadFromStackSlot
should return an array of accesses, rather than just the first encounter
of such an access.
This better describes FI accesses for AArch64 (paired) LDP/STP
instructions.
Reviewers: t.p.northover, gberry, thegameg, rengolin, javed.absar, MatzeB
Reviewed By: MatzeB
Differential Revision: https://reviews.llvm.org/D51537
llvm-svn: 341301
These intrinsics use the same implementation as PTEST intrinsics, but use vXi1 vectors.
New clang builtins will be accompanying them shortly.
llvm-svn: 341259
This patch recognizes shuffles that shift elements and fill with zeros. I've copied and modified the shift matching code we use for normal vector registers to do this. I'm not sure if there's a good way to share more of this code without making the existing function more complex than it already is.
This will be used to enable kshift intrinsics in clang.
Differential Revision: https://reviews.llvm.org/D51401
llvm-svn: 341227
The presence of a ReadAdvance for input operand #0 is problematic
because it changes the input latency of the register used as the base address
for the folded load.
A broadcast cannot start executing if the load address hasn't been computed yet.
In the llvm-mca example, the VBROADCASTSS is dependent on the address generated
by the LEAQ. That means, it cannot start until LEAQ reaches the write-back
stage. If we apply ReadAdvance, then we wrongly assume that the load can start 3
cycles in advance.
Differential Revision: https://reviews.llvm.org/D51534
llvm-svn: 341222
This patch fixes the number of micro opcodes, and processor resource cycles for
the following AVX instructions:
vinsertf128rr/rm
vperm2f128rr/rm
vbroadcastf128
Tests have been regenerated using the usual scripts in the llvm/utils directory.
Differential Revision: https://reviews.llvm.org/D51492
llvm-svn: 341185
These stubs should never be emitted for internal symbols, and
nothing in AsmPrinter ever actually use this value when producing
the stubs for COFF anyway.
llvm-svn: 341177
This assert tried to check that AND constants are only on the RHS. But its possible for both operands to be constants if one is opaque which will prevent the AND from being constant folded.
Fixes PR38771
llvm-svn: 341102
..Move all target-dependent checks into new isCopyInstrImpl method.
This change allows us to treat MoveReg-type instructions and generic
COPY instruction in the same way
Differential Revision: https://reviews.llvm.org/D49913
llvm-svn: 341072
We now only add +64bit to the CPU string for "generic" CPU. All other CPU names are assumed to have the feature flag already set if they support 64-bit. I've remove the implies from CMPXCHG8 so that Feature64Bit only comes in via CPUs or user passing -mattr=+64bit.
I've changed the assert to a report_fatal_error so it's not lost in Release builds.
The test updates are to fix things that tripped the new error.
Differential Revision: https://reviews.llvm.org/D51231
llvm-svn: 341022
Variables declared with the dllimport attribute are accessed via a
stub variable named __imp_<var>. In MinGW configurations, variables that
aren't declared with a dllimport attribute might still end up imported
from another DLL with runtime pseudo relocs.
For x86_64, this avoids the risk that the target is out of range
for a 32 bit PC relative reference, in case the target DLL is loaded
further than 4 GB from the reference. It also avoids having to make the
text section writable at runtime when doing the runtime fixups, which
makes it worthwhile to do for i386 as well.
Add stub variables for all dso local data references where a definition
of the variable isn't visible within the module, since the DLL data
autoimporting might make them imported even though they are marked as
dso local within LLVM.
Don't do this for variables that actually are defined within the same
module, since we then know for sure that it actually is dso local.
Don't do this for references to functions, since there's no need for
runtime pseudo relocations for autoimporting them; if a function from
a different DLL is called without the appropriate dllimport attribute,
the call just gets routed via a thunk instead.
GCC does something similar since 4.9 (when compiling with -mcmodel=medium
or large; from that version, medium is the default code model for x86_64
mingw), but only for x86_64.
Differential Revision: https://reviews.llvm.org/D51288
llvm-svn: 340942
Noticed while looking at D49562 codegen - we can avoid a large constant mask load and a slow VPBLENDVB select op by using VPBLENDW+VPBLENDD instead.
TODO: As discussed on the patch, we should investigate adding VPBLENDVB handling to target shuffle combining as well, that will allow us to extend this to VPBLENDW+VPBLENDW+VPBLENDD.
Differential Revision: https://reviews.llvm.org/D50074
llvm-svn: 340913
These are intrinsics for supporting kadd builtins in clang. These builtins are already in gcc to implement intrinsics from icc. Though they are missing from the Intel Intrinsics Guide.
This instruction adds two mask registers together as if they were scalar rather than a vXi1. We might be able to get away with a bitcast to scalar and a normal add instruction, but that would require DAG combine smarts in the backend to recoqnize add+bitcast. For now I'd prefer to go with the easiest implementation so we can get these builtins in to clang with good codegen.
Differential Revision: https://reviews.llvm.org/D51370
llvm-svn: 340869
These instructions were added on the PentiumPro along with CMOV.
This was already comprehended by the lowering process which should emit an alternate sequence using FCOM and FNSTW. This just makes it an explicit error if that doesn't work for some reason.
llvm-svn: 340844
This patch creates the shift mask and actual shift using the vXi16 vector shift ops.
Differential Revision: https://reviews.llvm.org/D51263
llvm-svn: 340813
We're using a 256-bit PACKUS to do the truncation, but that instruction operates on 128-bit lanes. So previously we shuffled first to rearrange the lanes. But that requires 2 shuffles. Instead we can shuffle after the PACKUS using a single VPERMQ. This matches what our normal LowerTRUNCATE code does when it uses PACKUS.
Differential Revision: https://reviews.llvm.org/D51284
llvm-svn: 340757
InstCombine mucks these up a bit. So we need to do some additional pattern matching to fix it. There are a still a few special cases not handled, but this covers the general case.
Differential Revision: https://reviews.llvm.org/D50952
llvm-svn: 340756
vXi32 support was recently moved from LowerMUL_LOHI to LowerMULH.
This commit shares the getOperand calls, switches both to use common IsSigned flag, and hoists the NumElems/NumElts variable.
llvm-svn: 340720
Summary: This was inheriting the cost from the AVX table, but should be legal under AVX512.
Reviewers: RKSimon
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51267
llvm-svn: 340708
Summary:
Previously most CPUs inherited cmov support through Feature64Bit(or FeatureCMPXCHG16HB implying Feature64Bit) or FeatureSSE1.
This has the surprising side effect that -mattr=-cmov causes an assert to fire in 64-bit mode because it clears the Feature64Bit. Or in 32-bit mode, -mattr=-cmov disables any sse/avx features which seems surprising.
This patch removes the implication and instead updates hasCMOV in X86Subtarget to check SSE1 or is64Bit in addition to the regular cmov flag. This should keep most things working the way they did before. I don't believe there is a way to specific "-cmov" directly from clang so this should only effect our lower level tools.
This does stop -mattr=cx16(cmpxchg16b) from implying cmov is enabled via the 64bit flag as you can see from one of the changed tests. But that was a 32-bit test so I don't know why it enabled cx16 anyway.
For the other test I had to add -sse to override the new sse check in hasCMOV.
Reviewers: RKSimon, DavidKreitzer, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits, jfb
Differential Revision: https://reviews.llvm.org/D51228
llvm-svn: 340707
Summary: This matches gcc and one cpuid dump I found online. Given that these are considered 7th generation x86 CPU it seems likely they support cmov since cmov was added by Intel in their 6th generation.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51264
llvm-svn: 340706
I noticed this along with the patterns in D51125, but when the index is variable,
we don't convert insertelement into a build_vector.
For x86, that means these get expanded at legalization time into the loading/spilling
code that we see in the tests. I think it's always better to avoid going to memory on
these, and we get the optimal 'broadcast' if it's available.
I suspect other targets may want to look at enabling the hook. AArch64 and AMDGPU have
regression tests that would be affected (although I did not check what would happen in
those cases). In the most basic cases shown here, AArch64 would probably do much
better with a splat.
Differential Revision: https://reviews.llvm.org/D51186
llvm-svn: 340705
Summary:
The only time vector SMUL_LOHI/UMUL_LOHI nodes are created is during division/remainder lowering. If its created before op legalization, generic DAGCombine immediately turns that SMUL_LOHI/UMUL_LOHI into a MULHS/MULHU since only the upper half is used. That node will stick around through vector op legalization and will be turned back into UMUL_LOHI/SMUL_LOHI during op legalization. It will then be custom lowered by the X86 backend. Due to this two step lowering the vector shuffles created by the custom lowering get legalized after their inputs rather than before. This prevents the shuffles from being combined with any build_vector of constants.
This patch uses changes vXi32 to use MULHS/MULHU instead. This is what the later DAG combine did anyway. But by skipping the change back to UMUL_LOHI/SMUL_LOHI we lower it before any constant BUILD_VECTORS. This allows the vector_shuffle creation to constant fold with the build_vectors. This accounts for the test changes here.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51254
llvm-svn: 340690
Summary:
Previously the value being stored is the last operand in SDNode. This causes the type legalizer to visit the mask operand before the value operand. The type legalizer was more complicated because of this since we want the type of the value to drive the decisions.
This patch moves the value to be the first operand so we visit it first during type legalization. It also simplifies the type legalization code accordingly.
X86 is currently the only in tree target that uses this SDNode. Not sure if there are any users out of tree.
Reviewers: RKSimon, delena, hfinkel, eli.friedman
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D50402
llvm-svn: 340689
This is a preliminary step for a preliminary step for D50992.
I noticed that x86 often misses chances to load a scalar directly
into a vector register.
So this patch is just allowing more of those cases to match a
broadcast op in lowerBuildVectorAsBroadcast(). The old code comment
said it doesn't make sense to use a broadcast when we're loading a
single element and everything else is undef, but I think that's the
best case in the improved tests in insert-loaded-scalar.ll. We avoid
scalar-to-vector-register move and/or less efficient shuffling.
Note that there are some existing types that were already producing
a broadcast, but that happens semi-accidentally. Ie, it's not
happening as part of lowerBuildVectorAsBroadcast(). The build vector
gets expanded into load + shuffle, and then shuffle lowering produces
the broadcast.
Description of the other test diffs:
1. avx-basic.ll - replacing load+shufle is a win.
2. sse3-avx-addsub-2.ll - vmovddup vs. vbroadcastss is neutral
3. sse41.ll - don't care - we convert that intrinsic to generic IR now, so this test is deprecated
4. vector-shuffle-128-v8.ll / vector-shuffle-256-v16.ll - pshufb alternatives with an extra instruction are not obviously bad
Differential Revision: https://reviews.llvm.org/D51125
llvm-svn: 340685
This adds a new method to ELFObjectFileBase that returns the symbols and addresses of PLT entries.
This design was suggested by pcc and eugenis in https://reviews.llvm.org/D49383.
Differential Revision: https://reviews.llvm.org/D50203
llvm-svn: 340610
The commit that added this functionality:
rL322957
may be causing/exposing a miscompile in PR38648:
https://bugs.llvm.org/show_bug.cgi?id=38648
so allow enabling/disabling to make debugging easier.
llvm-svn: 340540
subtarget features for indirect calls and indirect branches.
This is in preparation for enabling *only* the call retpolines when
using speculative load hardening.
I've continued to use subtarget features for now as they continue to
seem the best fit given the lack of other retpoline like constructs so
far.
The LLVM side is pretty simple. I'd like to eventually get rid of the
old feature, but not sure what backwards compatibility issues that will
cause.
This does remove the "implies" from requesting an external thunk. This
always seemed somewhat questionable and is now clearly not desirable --
you specify a thunk the same way no matter which set of things are
getting retpolines.
I really want to keep this nicely isolated from end users and just an
LLVM implementation detail, so I've moved the `-mretpoline` flag in
Clang to no longer rely on a specific subtarget feature by that name and
instead to be directly handled. In some ways this is simpler, but in
order to preserve existing behavior I've had to add some fallback code
so that users who relied on merely passing -mretpoline-external-thunk
continue to get the same behavior. We should eventually remove this
I suspect (we have never tested that it works!) but I've not done that
in this patch.
Differential Revision: https://reviews.llvm.org/D51150
llvm-svn: 340515
Previously we asumed a vector reduction add is part of a loop and one of the input is a phi. But the code in SelectionDAGBuilder that sets vector reduction flag handles more cases than that. It just requires that the use chain ends in a horizontal reduction. And there are no other uses. This means it can handle unrolled reduction loops.
If the initial value of the reduction was 0, an unrolled loop would begin with a vector reduction add that has two sad inputs. Previously we would only transform one side of the add, but for this case we need to transform both sides.
I've created a lambda to reuse some of the code for both sides. And fixed the variables names to remove reference to "phi".
Differential Revision: https://reviews.llvm.org/D50817
llvm-svn: 340478
Inspired by what AArch64 does for shifts, this patch attempts to replace shift amounts with neg if we can.
This is done directly as part of isel so its as late as possible to avoid breaking some BZHI patterns since those patterns need an unmasked (32-n) to be correct.
To avoid manual load folding and custom instruction selection for the negate. I've inserted new nodes in the DAG above the shift node in topological order.
Differential Revision: https://reviews.llvm.org/D48789
llvm-svn: 340441
When the key is not already in the map, the access operator[] creates an empty value and grows the map.
Resizing a map is very slow, so this needs to be avoided.
Found with csmith + asserts.
May help with
https://bugs.llvm.org/show_bug.cgi?id=25843
Patch by Tom Rix.
Differential Revision: https://reviews.llvm.org/D50780
llvm-svn: 340434
This adds the plumbing for the Tiny code model for the AArch64 backend. This,
instead of loading addresses through the normal ADRP;ADD pair used in the Small
model, uses a single ADR. The 21 bit range of an ADR means that the code and
its statically defined symbols need to be within 1MB of each other.
This makes it mostly interesting for embedded applications where we want to fit
as much as we can in as small a space as possible.
Differential Revision: https://reviews.llvm.org/D49673
llvm-svn: 340397
Summary:
So far, `isReturn` property is used to mean both a return instruction
from a functon and the end of an EH scope, a scope that starts with a EH
scope entry BB and ends with a catchret or a cleanupret instruction.
Because WinEH uses funclets, all EH-scope-ending instructions are also
real return instruction from a function. But for wasm, they only serve
as the end marker of an EH scope but not a return instruction that
exits a function. This mismatch caused incorrect prolog and epilog
generation in wasm EH scopes. This patch fixes this.
This patch is in the same vein with rL333045, which splits
`MachineBasicBlock::isEHFuncletEntry` into `isEHFuncletEntry` and
`isEHScopeEntry`.
Reviewers: dschuff
Subscribers: sbc100, jgravelle-google, sunfish, llvm-commits
Differential Revision: https://reviews.llvm.org/D50653
llvm-svn: 340325
Most of these shifts are extended to vXi16 so we don't gain anything from forcing another round of generic shift lowering - we know these extended cases are legal constant splat shifts.
llvm-svn: 340307
Due to some splat handling code in getVectorShuffle, its possible for NewV1/NewV2 to have their mask modified from what is requested. This can lead to cycles being created in the DAG.
This patch examines the returned mask and makes sure its different. Long term we may need to look closer at that splat code in getVectorShuffle, or add more splat awareness to getVectorShuffle.
Fixes PR38639
Differential Revision: https://reviews.llvm.org/D50981
llvm-svn: 340214
We can safely avoid interfering with the subus combine if both inputs are freely truncatable. Either both extends, or an extend and a constant vector.
Differential Revision: https://reviews.llvm.org/D50878
llvm-svn: 340212
We were basically assuming only one operand of the compare could be an ADD node and using that to swap operands. But we can have a normal add followed by a saturing add.
This rewrites the canonicalization to just be based on the condition code.
llvm-svn: 340134
The code already support 128 and 256 and even knows to split 256 for AVX1. So we really just needed to stop looking for specific VTs and subtarget features and just look for legal VTs with i8/i16 elements.
While there, add some curly braces around outer if statement bodies that contain only another if. It makes all the closing curly braces look more regular.
llvm-svn: 340128
Extending the concept introduced in D49562, this patch lowers constant vXi8 ISD::SRL/ISD::SRA by zero/sign extending to vXi16 and using PMULLW and then truncating the high 8 bits of the result.
Differential Revision: https://reviews.llvm.org/D50781
llvm-svn: 340062
isOnlyUserOf is a little heavier because it allows the node to be used multiple times by the other node. In this case we are looking at a truncate which only has one operand so we know it can only use it once. Thus hasOneUse is better.
llvm-svn: 340059
test/CodeGen/X86/shadow-stack.ll has the following machine verifier
errors:
```
*** Bad machine code: Using a killed virtual register ***
- function: bar
- basic block: %bb.6 entry (0x7fdc81857818)
- instruction: %3:gr64 = MOV64rm killed %2:gr64, 1, $noreg, 8, $noreg
- operand 1: killed %2:gr64
*** Bad machine code: Using a killed virtual register ***
- function: bar
- basic block: %bb.6 entry (0x7fdc81857818)
- instruction: $rsp = MOV64rm killed %2:gr64, 1, $noreg, 16, $noreg
- operand 1: killed %2:gr64
*** Bad machine code: Virtual register killed in block, but needed live out. ***
- function: bar
- basic block: %bb.2 entry (0x7fdc818574f8)
Virtual register %2 is used after the block.
```
The fix here is to only copy the machine operand's register without the
kill flags for all the instructions except the very last one of the
sequence.
I had to insert dummy PHIs in the test case to force the NoPHI function
property to be set to false. More on this here: https://llvm.org/PR38439
Differential Revision: https://reviews.llvm.org/D50260
llvm-svn: 340033
Normally the peephole pass converts EXTRACT_SUBREG to COPY instructions. But we're after peephole so we can't rely on it to clean these up.
To fix this, the eflags pass now emits a COPY with a subreg input.
I also noticed that in 32-bit mode we need to constrain the input to the copy to ensure the subreg is valid. Otherwise we'll fail verify-machineinstrs
Differential Revision: https://reviews.llvm.org/D50656
llvm-svn: 339945
a generically extensible collection of extra info attached to
a `MachineInstr`.
The primary change here is cleaning up the APIs used for setting and
manipulating the `MachineMemOperand` pointer arrays so chat we can
change how they are allocated.
Then we introduce an extra info object that using the trailing object
pattern to attach some number of MMOs but also other extra info. The
design of this is specifically so that this extra info has a fixed
necessary cost (the header tracking what extra info is included) and
everything else can be tail allocated. This pattern works especially
well with a `BumpPtrAllocator` which we use here.
I've also added the basic scaffolding for putting interesting pointers
into this, namely pre- and post-instruction symbols. These aren't used
anywhere yet, they're just there to ensure I've actually gotten the data
structure types correct. I'll flesh out support for these in
a subsequent patch (MIR dumping, parsing, the works).
Finally, I've included an optimization where we store any single pointer
inline in the `MachineInstr` to avoid the allocation overhead. This is
expected to be the overwhelmingly most common case and so should avoid
any memory usage growth due to slightly less clever / dense allocation
when dealing with >1 MMO. This did require several ergonomic
improvements to the `PointerSumType` to reasonably support the various
usage models.
This also has a side effect of freeing up 8 bits within the
`MachineInstr` which could be repurposed for something else.
The suggested direction here came largely from Hal Finkel. I hope it was
worth it. ;] It does hopefully clear a path for subsequent extensions
w/o nearly as much leg work. Lots of thanks to Reid and Justin for
careful reviews and ideas about how to do all of this.
Differential Revision: https://reviews.llvm.org/D50701
llvm-svn: 339940
Summary:
This prefix was added in r333421, and it changed our dumper output to
say things like "CVRegEAX" instead of just "EAX". That's a functional
change that I'd rather avoid.
I tested GCC, Clang, and MSVC, and all of them support #pragma
push_macro. They don't issue warnings whem the macro is not defined
either.
I don't have a Mac so I can't test the real termios.h header, but I
looked at the termios.h sources online and looked for other conflicts.
I saw only the CR* macros, so those are the ones we work around.
Reviewers: zturner, JDevlieghere
Subscribers: hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D50851
llvm-svn: 339907
Allow the comparison of x86 registers in the evaluation of assembler
directives. This generalizes and simplifies the extension from r334022
to catch another case found in the Linux kernel.
Reviewers: rnk, void
Reviewed By: rnk
Subscribers: hiraditya, nickdesaulniers, llvm-commits
Differential Revision: https://reviews.llvm.org/D50795
llvm-svn: 339895
When compiling with /arch:AVX512 and optimizations turned on,
we could crash while emitting debug info because we did not
have CodeView register constants for the AVX 512 register
set defined. This patch defines them.
Differential Revision: https://reviews.llvm.org/D50819
llvm-svn: 339893
a shorter name ('x86-slh') for the internal flags and pass name.
Without this, you can't use the -stop-after or -stop-before
infrastructure. I seem to have just missed this when originally adding
the pass.
The shorter name solves two problems. First, the flag names were ...
really long and hard to type/manage. Second, the pass name can't be the
exact same as the flag name used to enable this, and there are already
some users of that flag name so I'm avoiding changing it unnecessarily.
llvm-svn: 339836
To lower this we now create a new V1 containing the low half of both sources and a new V2 containing the upper half of both sources. Then we created a repeated lane shuffle of those new sources to create the final result.
This fixes PR35833
Differential Revison: https://reviews.llvm.org/D41794
llvm-svn: 339818
AVX512 added new versions of these intrinsics that take a rounding mode. If the rounding mode is 4 the new intrinsics are equivalent to the old intrinsics.
The AVX512 intrinsics were being lowered to ISD opcodes, but the legacy SSE intrinsics were left as intrinsics. This resulted in the AVX512 instructions needing separate patterns for the ISD opcodes and the legacy SSE intrinsics.
Now we convert SSE intrinsics and AVX512 intrinsics with rounding mode 4 to the same ISD opcode so we can share the isel patterns.
llvm-svn: 339749
`MachineMemOperand` pointers attached to `MachineSDNodes` and instead
have the `SelectionDAG` fully manage the memory for this array.
Prior to this change, the memory management was deeply confusing here --
The way the MI was built relied on the `SelectionDAG` allocating memory
for these arrays of pointers using the `MachineFunction`'s allocator so
that the raw pointer to the array could be blindly copied into an
eventual `MachineInstr`. This creates a hard coupling between how
`MachineInstr`s allocate their array of `MachineMemOperand` pointers and
how the `MachineSDNode` does.
This change is motivated in large part by a change I am making to how
`MachineFunction` allocates these pointers, but it seems like a layering
improvement as well.
This would run the risk of increasing allocations overall, but I've
implemented an optimization that should avoid that by storing a single
`MachineMemOperand` pointer directly instead of allocating anything.
This is expected to be a net win because the vast majority of uses of
these only need a single pointer.
As a side-effect, this makes the API for updating a `MachineSDNode` and
a `MachineInstr` reasonably different which seems nice to avoid
unexpected coupling of these two layers. We can map between them, but we
shouldn't be *surprised* at where that occurs. =]
Differential Revision: https://reviews.llvm.org/D50680
llvm-svn: 339740
This patch removes redundant template argument `TargetName` from TIIPredicate.
Tablegen can always infer the target name from the context. So we don't need to
force users of TIIPredicate to always specify it.
This allows us to better modularize the tablegen class hierarchy for the
so-called "function predicates". class FunctionPredicateBase has been added; it
is currently used as a building block for TIIPredicates. However, I plan to
reuse that class to model other function predicate classes too (i.e. not just
TIIPredicates). For example, this can be a first step towards implementing
proper support for dependency breaking instructions in tablegen.
This patch also adds a verification step on TIIPredicates in tablegen.
We cannot have multiple TIIPredicates with the same name. Otherwise, this will
cause build errors later on, when tablegen'd .inc files are included by cpp
files and then compiled.
Differential Revision: https://reviews.llvm.org/D50708
llvm-svn: 339706
rL339686 added the case where a faux shuffle might have repeated shuffle inputs coming from either side of the OR().
This patch improves the insertion of the inputs into the source ops lists to account for this, as well as making it trivial to add support for shuffles with more than 2 inputs in the future.
llvm-svn: 339696
Summary: This revision improves previous version (rL330322) which has been reverted due to crashes.
This is the patch that lowers x86 intrinsics to native IR
in order to enable optimizations. The patch also includes folding
of previously missing saturation patterns so that IR emits the same
machine instructions as the intrinsics.
Reviewers: craig.topper, spatel, RKSimon
Reviewed By: craig.topper
Subscribers: mike.dvoretsky, DavidKreitzer, sroland, llvm-commits
Differential Revision: https://reviews.llvm.org/D46179
llvm-svn: 339650
The behavior in 64-bit mode is different between Intel and AMD CPUs. Intel ignores the 0x66 prefix. AMD does not. objump doesn't ignore the 0x66 prefix. Since LLVM aims to match objdump behavior, we should do the same.
While I was trying to fix this I had change brtarget16/32 to use ENCODING_IW/ID instead of ENCODING_Iv to get the 0x66+REX.W case to act sort of sanely. It's still wrong, but that's a problem for another day.
The change in encoding exposed the fact that 16-bit mode disassembly of relative jumps was creating JMP_4 with a 2 byte immediate. It should have been JMP_2. From just printing you can't tell the difference, but if you dumped the encoding it wouldn't have matched what we started with.
While fixing that, it exposed that jo/jno opcodes were missing from the switch that this patch deleted and there were no test cases for them.
Fixes PR38537.
llvm-svn: 339622
Unlike the other arithmetic instructions the mem-reg form of compare is just a load and not a RMW operation. According to the Intel optimization manual, this form is also supported by macro fusion.
llvm-svn: 339498
Now we switch to the subregister in expandPostRAPseudos where we already switched the opcode.
This simplifies a few isel patterns that used the pseudo directly. And magically seems to have improved our ability to CSE it in the undef-label.ll test.
llvm-svn: 339496
This patch introduces tablegen class MCStatement.
Currently, an MCStatement can be either a return statement, or a switch
statement.
```
MCStatement:
MCReturnStatement
MCOpcodeSwitchStatement
```
A MCReturnStatement expands to a return statement, and the boolean expression
associated with the return statement is described by a MCInstPredicate.
An MCOpcodeSwitchStatement is a switch statement where the condition is a check
on the machine opcode. It allows the definition of multiple checks, as well as a
default case. More details on the grammar implemented by these two new
constructs can be found in the diff for TargetInstrPredicates.td.
This patch makes it easier to read the body of auto-generated TargetInstrInfo
predicates.
In future, I plan to reuse/extend the MCStatement grammar to describe more
complex target hooks. For now, this is just a first step (mostly a minor
cosmetic change to polish the new predicates framework).
Differential Revision: https://reviews.llvm.org/D50457
llvm-svn: 339352
As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses.
This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool.
However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move.
This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit.
Differential Revision: https://reviews.llvm.org/D50328
llvm-svn: 339335
Src0 doesn't really convey any meaning to what the operand is. Passthru matches what's used in the documentation for the intrinsic this comes from.
llvm-svn: 339101
Summary:
Expand isFNEG so that we generate the appropriate F(N)M(ADD|SUB)
instructions in more cases. For example, the following sequence
a = _mm256_broadcast_ss(f)
d = _mm256_fnmadd_ps(a, b, c)
generates an fsub and fma without this patch and an fnma with this
change.
Reviewers: craig.topper
Subscribers: llvm-commits, davidxl, wmi
Differential Revision: https://reviews.llvm.org/D48467
llvm-svn: 339043
If the store is volatile this might be a memory mapped IO access. In that case we shouldn't generate a load that didn't exist in the source
Differential Revision: https://reviews.llvm.org/D50270
llvm-svn: 339041
At one point in time acquire implied mayLoad and mayStore as did release. Thus we needed separate pseudos that also carried that property. This appears to no longer be the case. I believe it was changed in 2012 with a comment saying that atomic memory accesses are marked volatile which preserves the ordering.
So from what I can tell we shouldn't need additional pseudos since they aren't carry any flags that are different from the normal instructions. The only thing I can think of is that we may consider them for load folding candidates in the peephole pass now where we didn't before. If that's important hopefully there's something in the memory operand we can check to prevent the folding without relying on pseudo instructions.
Differential Revision: https://reviews.llvm.org/D50212
llvm-svn: 338925
Clang uses "ctpop & 1" to implement __builtin_parity. If the popcnt instruction isn't supported this generates a large amount of code to calculate the population count. Instead we can bisect the data down to a single byte using xor and then check the parity flag.
Even when popcnt is supported, its still a good idea to split 64-bit data on 32-bit targets using an xor in front of a single popcnt. Otherwise we get two popcnts and an add before the and.
I've specifically targeted this at the sizes supported by clang builtins, but we could generalize this if we think that's useful.
Differential Revision: https://reviews.llvm.org/D50165
llvm-svn: 338907
There are a lot of permutations of types here generating a lot of patterns in the isel table. It's more efficient to just ReplaceUses and RemoveDeadNode from the Select function.
The test changes are because we have a some shuffle patterns that have a bitcast as their root node. But the behavior is identical to another instruction whose pattern doesn't start with a bitcast. So this isn't a functional change.
llvm-svn: 338824
Move all the patterns to X86InstrVecCompiler.td so we can keep SSE/AVX/AVX512 all in one place.
To save some patterns we'll use an existing DAG combine to convert f128 fand/for/fxor to integer when sse2 is enabled. This allows use to reuse all the existing patterns for v2i64.
I believe this now makes SHA instructions the only case where VEX/EVEX and legacy encoded instructions could be generated simultaneously.
llvm-svn: 338821
If the producing instruction is legacy encoded it doesn't implicitly zero the upper bits. This is important for the SHA instructions which don't have a VEX encoded version. We might also be able to hit this with the incomplete f128 support that hasn't been ported to VEX.
llvm-svn: 338812
I'm assuming the R13 restriction extends to R13D. Guessing this restriction is related to the funny encoding of this register as base always requiring a displacement to be encoded.
llvm-svn: 338806
These instructions perform the same operation, but the semantic of which operand is destroyed is reversed. If the same register is used as both operands we can change the execution domain without worrying about this difference.
Unfortunately, this really only works in cases where the input register is killed by the instruction. If its not killed, the two address isntruction pass inserts a copy that will become a move instruction. This makes the instruction use different physical registers that contain the same data at the time the unpck/movhlps executes. I've considered using a unary pseudo instruction with tied operand to trick the two address instruction pass. We could then expand the pseudo post regalloc to get the same physical register on both inputs.
Differential Revision: https://reviews.llvm.org/D50157
llvm-svn: 338735
We now emit a move of -1 before the cmov and do the addition after the cmov just like the case with an extra addition.
This may be slightly worse for code size, but is more consistent with other compilers. And we might be able to hoist the mov -1 outside of loops.
llvm-svn: 338613
Summary:
D25878, which added support for !absolute_symbol for normal X86 ISel,
did not add support for materializing references to absolute symbols for
X86 FastISel. This causes build failures because FastISel generates
PC-relative relocations for absolute symbols. Fall back to normal ISel
for references to !absolute_symbol GVs. Fix for PR38200.
Reviewers: pcc, craig.topper
Reviewed By: pcc
Subscribers: hiraditya, llvm-commits, kcc
Differential Revision: https://reviews.llvm.org/D50116
llvm-svn: 338599
There is nothing x86-specific about this code, so it'd be nice to make this available for other targets to use in the future (and get it out of X86ISelLowering!).
Differential Revision: https://reviews.llvm.org/D50083
llvm-svn: 338586
It's not strictly required by the transform of the cmov and the add, but it makes sure we restrict it to the cases we know we want to match.
While there canonicalize the operand order of the cmov to simplify the matching and emitting code.
llvm-svn: 338492
EFLAGS copy lowering.
If you have a branch of LLVM, you may want to cherrypick this. It is
extremely unlikely to hit this case empirically, but it will likely
manifest as an "impossible" branch being taken somewhere, and will be
... very hard to debug.
Hitting this requires complex conditions living across complex control
flow combined with some interesting memory (non-stack) initialized with
the results of a comparison. Also, because you have to arrange for an
EFLAGS copy to be in *just* the right place, almost anything you do to
the code will hide the bug. I was unable to reduce anything remotely
resembling a "good" test case from the place where I hit it, and so
instead I have constructed synthetic MIR testing that directly exercises
the bug in question (as well as the good behavior for completeness).
The issue is that we would mistakenly assume any SETcc with a valid
condition and an initial operand that was a register and a virtual
register at that to be a register *defining* SETcc...
It isn't though....
This would in turn cause us to test some other bizarre register,
typically the base pointer of some memory. Now, testing this register
and using that to branch on doesn't make any sense. It even fails the
machine verifier (if you are running it) due to the wrong register
class. But it will make it through LLVM, assemble, and it *looks*
fine... But wow do you get a very unsual and surprising branch taken in
your actual code.
The fix is to actually check what kind of SETcc instruction we're
dealing with. Because there are a bunch of them, I just test the
may-store bit in the instruction. I've also added an assert for sanity
that ensure we are, in fact, *defining* the register operand. =D
llvm-svn: 338481
Don't declare them as X86SchedWritePair when the folded class will never be used.
Note: MOVBE (load/store endian conversion) instructions tend to have a very different behaviour to BSWAP.
llvm-svn: 338412
As was done for vector rotations, we can efficiently use ISD::MULHU for vXi8/vXi16 ISD::SRL lowering.
Shift-by-zero cases are still problematic (mainly on v32i8 due to extra AND/ANDN/OR or VPBLENDVB blend masks but v8i16/v16i16 aren't great either if PBLENDW fails) so I've limited this first patch to known non-zero cases if we can't easily use PBLENDW.
Differential Revision: https://reviews.llvm.org/D49562
llvm-svn: 338407
Summary:
Similar to D49636, but for PMADDUBSW. This instruction has the additional complexity that the addition of the two products saturates to 16-bits rather than wrapping around. And one operand is treated as signed and the other as unsigned.
A C example that triggers this pattern
```
static const int N = 128;
int8_t A[2*N];
uint8_t B[2*N];
int16_t C[N];
void foo() {
for (int i = 0; i != N; ++i)
C[i] = MIN(MAX((int16_t)A[2*i]*(int16_t)B[2*i] + (int16_t)A[2*i+1]*(int16_t)B[2*i+1], -32768), 32767);
}
```
Reviewers: RKSimon, spatel, zvi
Reviewed By: RKSimon, zvi
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D49829
llvm-svn: 338402
This commit fixes two issues with the liveness information after the
call:
1) The code always spills RCX and RDX if InProlog == true, which results
in an use of undefined phys reg.
2) FinalReg, JoinReg, RoundedReg, SizeReg are not added as live-ins to
the basic blocks that use them, therefore they are seen undefined.
https://llvm.org/PR38376
Differential Revision: https://reviews.llvm.org/D50020
llvm-svn: 338400
This patch teaches llvm-mca how to identify dependency breaking instructions on
btver2.
An example of dependency breaking instructions is the zero-idiom XOR (example:
`XOR %eax, %eax`), which always generates zero regardless of the actual value of
the input register operands.
Dependency breaking instructions don't have to wait on their input register
operands before executing. This is because the computation is not dependent on
the inputs.
Not all dependency breaking idioms are also zero-latency instructions. For
example, `CMPEQ %xmm1, %xmm1` is independent on
the value of XMM1, and it generates a vector of all-ones.
That instruction is not eliminated at register renaming stage, and its opcode is
issued to a pipeline for execution. So, the latency is not zero.
This patch adds a new method named isDependencyBreaking() to the MCInstrAnalysis
interface. That method takes as input an instruction (i.e. MCInst) and a
MCSubtargetInfo.
The default implementation of isDependencyBreaking() conservatively returns
false for all instructions. Targets may override the default behavior for
specific CPUs, and return a value which better matches the subtarget behavior.
In future, we should teach to Tablegen how to automatically generate the body of
isDependencyBreaking from scheduling predicate definitions. This would allow us
to expose the knowledge about dependency breaking instructions to the machine
schedulers (and, potentially, other codegen passes).
Differential Revision: https://reviews.llvm.org/D49310
llvm-svn: 338372
isFNEG was duplicating much of what was done by getTargetConstantBitsFromNode in its own calls to getTargetConstantFromNode.
Noticed while reviewing D48467.
llvm-svn: 338358
In one place we checked X86Subtarget.slowLEA() to decide if the pass should run. But to decide what the pass should we only check isSLM. This resulted in Goldmont going down the Bonnell path.
llvm-svn: 338342
The machine verifier asserts with:
Assertion failed: (isMBB() && "Wrong MachineOperand accessor"), function getMBB, file ../include/llvm/CodeGen/MachineOperand.h, line 542.
It calls analyzeBranch which tries to call getMBB if the opcode is
JMP_1, but in this case we do:
JMP_1 @OUTLINED_FUNCTION
I believe we have to use TAILJMPd64 instead of JMP_1 since JMP_1 is used
with brtarget8.
Differential Revision: https://reviews.llvm.org/D49299
llvm-svn: 338237
X86 normally requires immediates to be a signed 32-bit value which would exclude i64 0x80000000. But for add/sub we can negate the constant and use the opposite instruction.
llvm-svn: 338204
Not sure why they were being explicitly excluded, but I believe all the math inside the if works. I changed the absolute value to be uint64_t instead of int64_t so INT64_MIN+1 wouldn't be signed wrap.
llvm-svn: 338101
Summary:
This is the pattern you get from the loop vectorizer for something like this
int16_t A[1024];
int16_t B[1024];
int32_t C[512];
void pmaddwd() {
for (int i = 0; i != 512; ++i)
C[i] = (A[2*i]*B[2*i]) + (A[2*i+1]*B[2*i+1]);
}
In this case we will have (add (mul (build_vector), (build_vector)), (mul (build_vector), (build_vector))). This is different than the pattern we currently match which has the build_vectors between an add and a single multiply. I'm not sure what C code would get you that pattern.
Reviewers: RKSimon, spatel, zvi
Reviewed By: zvi
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D49636
llvm-svn: 338097
If this happens the operands aren't updated and the existing node is returned. Make sure we pass this existing node up to the DAG combiner so that a proper replacement happens. Otherwise we get stuck in an infinite loop with an unoptimized node.
llvm-svn: 338090
a helper function with a nice overview comment. NFC.
This is a preperatory refactoring to implementing another component of
mitigation here that was descibed in the design document but hadn't been
implemented yet.
llvm-svn: 338016
I'm not sure if this was trying to avoid optimizing the new nodes further or what. Or maybe to prevent a cycle if something tried to reform the multiply? But I don't think its a reliable way to do that. If the user of the expanded multiply is visited by the DAGCombiner after this conversion happens, the DAGCombiner will check its operands, see that they haven't been visited by the DAGCombiner before and it will then add the first node to the worklist. This process will repeat until all the new nodes are visited.
So this seems like an unreliable prevention at best. So this patch just returns the new nodes like any other combine. If this starts causing problems we can try to add target specific nodes or something to more directly prevent optimizations.
Now that we handle the combine normally, we can combine any negates the mul expansion creates into their users since those will be visited now.
llvm-svn: 338007
These calls were making sure some newly created nodes were added to worklist, but the DAGCombiner has internal support for ensuring it has visited all nodes. Any time it visits a node it ensures the operands have been queued to be visited as well. This means if we only need to return the last new node. The DAGCombiner will take care of adding its inputs thus walking backwards through all the new nodes.
llvm-svn: 337996
- Avoid duplication of regmask size calculation.
- Simplify allocateRegisterMask() call.
- Rename allocateRegisterMask() to allocateRegMask() to be consistent
with naming in MachineOperand.
llvm-svn: 337986
In SVN r334523, the first half of comdat constant pool handling was
hoisted from X86WindowsTargetObjectFile (which despite the name only
was used for msvc targets) into the arch independent
TargetLoweringObjectFileCOFF, but the other half of the handling was
left behind in X86AsmPrinter::GetCPISymbol.
With only half of the handling in place, inconsistent comdat
sections/symbols are created, causing issues with both GNU binutils
(avoided for X86 in SVN r335918) and with the MS linker, which
would complain like this:
fatal error LNK1143: invalid or corrupt file: no symbol for COMDAT section 0x4
Differential Revision: https://reviews.llvm.org/D49644
llvm-svn: 337950
code.
This consolidates all our hardening calls, and simplifies the code
a bit. It seems much more clear to handle all of these together.
No functionality changed here.
llvm-svn: 337895
This function actually does two things: it traces the predicate state
through each of the basic blocks in the function (as that isn't directly
handled by the SSA updater) *and* it hardens everything necessary in the
block as it goes. These need to be done together so that we have the
currently active predicate state to use at each point of the hardening.
However, this also made obvious that the flag to disable actual
hardening of loads was flawed -- it also disabled tracing the predicate
state across function calls within the body of each block. So this patch
sinks this debugging flag test to correctly guard just the hardening of
loads.
Unless load hardening was disabled, no functionality should change with
tis patch.
llvm-svn: 337894
against v1.2 BCBS attacks directly.
Attacks using spectre v1.2 (a subset of BCBS) are described in the paper
here:
https://people.csail.mit.edu/vlk/spectre11.pdf
The core idea is to speculatively store over the address in a vtable,
jumptable, or other target of indirect control flow that will be
subsequently loaded. Speculative execution after such a store can
forward the stored value to subsequent loads, and if called or jumped
to, the speculative execution will be steered to this potentially
attacker controlled address.
Up until now, this could be mitigated by enableing retpolines. However,
that is a relatively expensive technique to mitigate this particular
flavor. Especially because in most cases SLH will have already mitigated
this. To fully mitigate this with SLH, we need to do two core things:
1) Unfold loads from calls and jumps, allowing the loads to be post-load
hardened.
2) Force hardening of incoming registers even if we didn't end up
needing to harden the load itself.
The reason we need to do these two things is because hardening calls and
jumps from this particular variant is importantly different from
hardening against leak of secret data. Because the "bad" data here isn't
a secret, but in fact speculatively stored by the attacker, it may be
loaded from any address, regardless of whether it is read-only memory,
mapped memory, or a "hardened" address. The only 100% effective way to
harden these instructions is to harden the their operand itself. But to
the extent possible, we'd like to take advantage of all the other
hardening going on, we just need a fallback in case none of that
happened to cover the particular input to the control transfer
instruction.
For users of SLH, currently they are paing 2% to 6% performance overhead
for retpolines, but this mechanism is expected to be substantially
cheaper. However, it is worth reminding folks that this does not
mitigate all of the things retpolines do -- most notably, variant #2 is
not in *any way* mitigated by this technique. So users of SLH may still
want to enable retpolines, and the implementation is carefuly designed to
gracefully leverage retpolines to avoid the need for further hardening
here when they are enabled.
Differential Revision: https://reviews.llvm.org/D49663
llvm-svn: 337878
We generated a subtract for the power of 2 minus one then negated the result. The negate can be optimized away by swapping the subtract operands, but DAG combine doesn't know how to do that and we don't add any of the new nodes to the worklist anyway.
This patch makes use explicitly emit the swapped subtract.
llvm-svn: 337858
Use a left shift and 2 subtracts like we do for 30. Move this out from behind the slow lea check since it doesn't even use an LEA.
Use this for multiply by 14 as well.
llvm-svn: 337856
Just some gardening here.
Similar to how we moved call information into Candidates, this moves outlined
frame information into OutlinedFunction. This allows us to remove
TargetCostInfo entirely.
Anywhere where we returned a TargetCostInfo struct, we now return an
OutlinedFunction. This establishes OutlinedFunctions as more of a general
repeated sequence, and Candidates as occurrences of those repeated sequences.
llvm-svn: 337848
Summary:
Enabling this fully exposes a latent bug in the instruction folding: we
never update the register constraints for the register operands when
fusing a load into another operation. The fused form could, in theory,
have different register constraints on its operands. And in fact,
TCRETURNm* needs its memory operands to use tailcall compatible
registers.
I've updated the folding code to re-constrain all the registers after
they are mapped onto their new instruction.
However, we still can't enable folding in the general case from
TCRETURNr* to TCRETURNm* because doing so may require more registers to
be available during the tail call. If the call itself uses all but one
register, and the folded load would require both a base and index
register, there will not be enough registers to allocate the tail call.
It would be better, IMO, to teach the register allocator to *unfold*
TCRETURNm* when it runs out of registers (or specifically check the
number of registers available during the TCRETURNr*) but I'm not going
to try and solve that for now. Instead, I've just blocked the forward
folding from r -> m, leaving LLVM free to unfold from m -> r as that
doesn't introduce new register pressure constraints.
The down side is that I don't have anything that will directly exercise
this. Instead, I will be immediately using this it my SLH patch. =/
Still worse, without allowing the TCRETURNr* -> TCRETURNm* fold, I don't
have any tests that demonstrate the failure to update the memory operand
register constraints. This patch still seems correct, but I'm nervous
about the degree of testing due to this.
Suggestions?
Reviewers: craig.topper
Subscribers: sanjoy, mcrosier, hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D49717
llvm-svn: 337845
Before this, TCI contained all the call information for each Candidate.
This moves that information onto the Candidates. As a result, each Candidate
can now supply how it ought to be called. Thus, Candidates will be able to,
say, call the same function in cheaper ways when possible. This also removes
that information from TCI, since it's no longer used there.
A follow-up patch for the AArch64 outliner will demonstrate this.
llvm-svn: 337840
helper and restructure the post-load hardening to use this.
This isn't as trivial as I would have liked because the post-load
hardening used a trick that only works for it where it swapped in
a temporary register to the load rather than replacing anything.
However, there is a simple way to do this without that trick that allows
this to easily reuse a friendly API for hardening a value in a register.
That API will in turn be usable in subsequent patcehs.
This also techincally changes the position at which we insert the subreg
extraction for the predicate state, but that never resulted in an actual
instruction and so tests don't change at all.
llvm-svn: 337825
This code was really nasty, had several bugs in it originally, and
wasn't carrying its weight. While on Zen we have all 4 ports available
for SHRX, on all of the Intel parts with Agner's tables, SHRX can only
execute on 2 ports, giving it 1/2 the throughput of OR.
Worse, all too often this pattern required two SHRX instructions in
a chain, hurting the critical path by a lot.
Even if we end up needing to safe/restore EFLAGS, that is no longer so
bad. We pay for a uop to save the flag, but we very likely get fusion
when it is used by forming a test/jCC pair or something similar. In
practice, I don't expect the SHRX to be a significant savings here, so
I'd like to avoid the complex code required. We can always resurrect
this if/when someone has a specific performance issue addressed by it.
llvm-svn: 337781
Don't try to generate large PIC code for non-ELF targets. Neither COFF
nor MachO have relocations for large position independent code, and
users have been using "large PIC" code models to JIT 64-bit code for a
while now. With this change, if they are generating ELF code, their
JITed code will truly be PIC, but if they target MachO or COFF, it will
contain 64-bit immediates that directly reference external symbols. For
a JIT, that's perfectly fine.
llvm-svn: 337740
Summary:
Pretty mechanical follow-up for D49196.
As microarchitecture.pdf notes, "20 AMD Ryzen pipeline",
"20.8 Register renaming and out-of-order schedulers":
The integer register file has 168 physical registers of 64 bits each.
The floating point register file has 160 registers of 128 bits each.
"20.14 Partial register access":
The processor always keeps the different parts of an integer register together.
...
An instruction that writes to part of a register will therefore have a false dependence
on any previous write to the same register or any part of it.
Reviewers: andreadb, courbet, RKSimon, craig.topper, GGanesh
Reviewed By: GGanesh
Subscribers: gbedwell, llvm-commits
Differential Revision: https://reviews.llvm.org/D49393
llvm-svn: 337676
a call, and then again as a return.
Also added a comment to try and explain better why we would be doing
what we're doing when hardening the (non-call) returns.
llvm-svn: 337673
This provides an overview of the algorithm used to harden specific
loads. It also brings this our terminology further in line with
hardening rather than checking.
Differential Revision: https://reviews.llvm.org/D49583
llvm-svn: 337667
This seems to be a net improvement. There's still an issue under avx512f where we have a 512-bit vpaddd, but not vpmaddwd so we end up doing two 256-bit vpmaddwds and inserting the results before a 512-bit vpaddd. It might be better to do two 512-bits paddds with zeros in the upper half. Same number of instructions, but breaks a dependency.
llvm-svn: 337656
Craig Topper